d-08-06[1]

The views expressed are purely those of the writers and may not, in any circumstances, be regarded as stating an official position of the European Commission

EUFIRELAB

EVR1-CT-2002-40028

D-08-06

http://eufirelab.org

EUFIRELAB: Euro-Mediterranean Wildland Fire Laboratory,

a “wall-less” Laboratory for Wildland Fire Sciences and Technologies

in the Euro-Mediterranean Region

Deliverable D-08-06

Towards a Euro-Mediterranean Wildland Fire Danger Rating System

AUTHORS by PARTNER

P014: Giovanni BOVIO, Raffaella MARZANO P023: Inmaculada AGUADO, Emilio CHUVIECO, Jesús MARTÍNEZ

Héctor NIETO, Javier SALAS P027: Wanda BEROLO, Pierre CARREGA, Dennis FOX, Nuno GERONIMO,

Jean-Guillaume ROBIN P030: Israel GÓMEZ, Pilar MARTÍN, Javier MARTÍNEZ-VEGA, Lara VILAR

P036: Ioannis GITAS, Michael KARTERIS, Spyros TSAKALIDIS

December 2006

EUFIRELAB

D-08-06.doc

CONTENT LIST

1 Scope and objectives...............................................................................................................................................1 1.1 First phase.....................................................................................................................................................1 1.2 Second phase................................................................................................................................................2 1.3 Third phase....................................................................................................................................................2

2 Proposed methods for wildland fire danger estimation at European scale .............................................................3 3 Review of fire danger rating System developed in MEGAFIRES project................................................................4

3.1 Introduction....................................................................................................................................................4 3.2 Use of meteorological indices and satellite data ...........................................................................................4 3.3 Figures...........................................................................................................................................................6 3.4 Tables............................................................................................................................................................8 3.5 Analysis of long-term fire risk on a European level .......................................................................................9

3.5.1 Selection of risk variables....................................................................................................................9 3.5.2 Techniques to estimate large fire occurrence ...................................................................................10 3.5.3 Figures...............................................................................................................................................12 3.5.4 Tables................................................................................................................................................14

4 Review of fire risk estimation method developed for the SPREAD project ...........................................................15 4.1 Input data for fire risk mapping....................................................................................................................16 4.2 Human danger ignition ................................................................................................................................16 4.3 Probability of Ignition and Fuel Moisture Content .......................................................................................17

4.3.1 Fuel Moisture Content of live fuels ....................................................................................................17 4.3.2 Fuel Moisture Content of dead fuels .................................................................................................17 4.3.3 Probability of Ignition related to the FMC ..........................................................................................17

4.4 Propagation Danger ....................................................................................................................................18 4.4.1 Average Rate of Spread and Flame Length......................................................................................18 4.4.2 Propagation danger versus RoS and FL...........................................................................................18

4.5 Wildland Fire Danger Assessment..............................................................................................................19 4.6 Figures.........................................................................................................................................................19

5 Review of the European Forest Fire Information System-Risk Forecast ..............................................................22 5.1 Introduction..................................................................................................................................................22 5.2 Meteorological danger indices.....................................................................................................................23 5.3 FPI Model. ...................................................................................................................................................23

5.3.1 General presentation.........................................................................................................................23 5.3.2 Input data...........................................................................................................................................24 5.3.3 Computation of FPI............................................................................................................................26

5.4 Long-term indices........................................................................................................................................26 5.4.1 Probability of Fire Occurrence...........................................................................................................26 5.4.2 Likely Damage...................................................................................................................................27

5.5 Figures.........................................................................................................................................................28 5.6 Tables..........................................................................................................................................................29

6 Basic structure and characteristics of the proposed Euro-Mediterranean Wildland Fire Risk Index ....................31 6.1 General presentation...................................................................................................................................31 6.2 Figures.........................................................................................................................................................33

7 Components of the fire risk system: Ignition Danger Index...................................................................................34 7.1 Fuel Moisture...............................................................................................................................................34

7.1.1 Introduction........................................................................................................................................34 7.1.2 Live fuels (satellite information).........................................................................................................34 7.1.3 Study case in Central Spain ..............................................................................................................35 7.1.4 Probability of ignition related to live fuel moisture content ................................................................39 7.1.5 Probability of Ignition in Dead Fuels (Meteorological Index).............................................................40 7.1.6 Figures...............................................................................................................................................44

7.2 Probability of Ignition (Human Factors).......................................................................................................49 7.2.1 Introduction........................................................................................................................................49 7.2.2 Estimation of human ignition danger at regional scale: the case of Alpes-Maritimes (France) ........50 7.2.3 The third model: statistical approach B. ............................................................................................53 7.2.4 Model comparison and conclusion. ...................................................................................................53 7.2.5 Study Area and Wildland Fire Database ...........................................................................................54 7.2.6 Methodology: Geographically Weighted Regression (GWR), the linear and logistic case ...............55 7.2.7 Results and discussion......................................................................................................................55

EUFIRELAB

D-08-06.doc

7.2.8 Conclusion.........................................................................................................................................56 7.2.9 Figures...............................................................................................................................................57 7.2.10 Tables................................................................................................................................................61

7.3 Human ignition danger in Southern Europe based on fire occurrence maps .............................................63 7.3.1 Fire Occurrence Hot Spot Areas .......................................................................................................63 7.3.2 Figures...............................................................................................................................................65 7.3.3 Tables................................................................................................................................................67

8 Propagation danger index .....................................................................................................................................68 8.1 Average rate of spread and Flame Length..................................................................................................68 8.2 Propagation danger (PD) ............................................................................................................................69 8.3 Figures.........................................................................................................................................................69

9 Vulnerability index..................................................................................................................................................70 9.1 Population vulnerability ...............................................................................................................................70

9.1.1 Introduction........................................................................................................................................70 9.1.2 Example of population vulnerability mapping in Piemonte Region (Italy) .........................................72 9.1.3 Estimation of population vulnerability at Euro-Mediterranean scale .................................................73 9.1.4 Figures...............................................................................................................................................76 9.1.5 Tables................................................................................................................................................80

9.2 Vulnerability related to environmental value ...............................................................................................82 9.2.1 Presentation ......................................................................................................................................82 9.2.2 Figures...............................................................................................................................................83

9.3 Potential soil erosion ...................................................................................................................................85 9.3.1 Introduction........................................................................................................................................85 9.3.2 Objectives..........................................................................................................................................85 9.3.3 A European scale model to predict potential erosion risk .................................................................86 9.3.4 An operational model for the Mediterranean context ........................................................................89 9.3.5 Figures...............................................................................................................................................92 9.3.6 Tables................................................................................................................................................99

10 Euro-Mediterranean Wildland Fire Risk Index.....................................................................................................100 10.1 General presentation.................................................................................................................................100 10.2 Figures.......................................................................................................................................................101

11 References ..........................................................................................................................................................102 12 Annex: Mapping post-fire soil erosion risk...........................................................................................................114

12.1 Introduction................................................................................................................................................114 12.2 The impact of forest fires on soil erodibility ...............................................................................................114 12.3 Site description..........................................................................................................................................115 12.4 Methods.....................................................................................................................................................116

12.4.1 Mapping soil erosion risk.................................................................................................................116 12.4.2 Soil erosion factors ..........................................................................................................................116 12.4.3 Partial validation of the soil erosion risk map ..................................................................................117

12.5 Results.......................................................................................................................................................117 12.5.1 Distribution of the soil erosion factors .............................................................................................117 12.5.2 The soil erosion risk map ................................................................................................................117 12.5.3 Partial model validation ...................................................................................................................118

12.6 Discussion .................................................................................................................................................118 12.7 Conclusions...............................................................................................................................................119 12.8 Figures.......................................................................................................................................................120 12.9 Tables........................................................................................................................................................122

EUFIRELAB

D-08-06.doc

SUMMARY

This deliverable is strongly related and in a sense the succession of the deliverables D-08-02 (Wildland fire danger and hazards: a state of the art), D-08-05 (Common methods for mapping the wildland fire danger), and D-08-03 (Towards a Euro-Mediterranean Wildland Fire Danger Rating System: basis, structure).

Deliverable D-08-03 addresses the design of the basis and structure of a prototype of a Euro-Mediterranean Wildland Fire Danger Rating System.

Deliverable D-08-06 addresses the final design of the system.

The present work has been divided in three main chapters.

In chapters 2 to 5, a review of methods already developed and applied operationally or semi-operationally for estimating and mapping wildland fire risks at Euro-Mediterranean scale is presented.

The results of two previous European projects (Megafires and Spread) and the European Forest Fire Information System-Risk Forecast –EFFIS- developed by the JRC are analysed.

In chapter 6 the basic structure of the wildland fire danger rating system is proposed.

This system is composed by the three most important components in forest fire risk: ignition, propagation and vulnerability.

The basis and structure have been defined taking into account the proposals and suggestion of the projects analysed in the previous chapter.

In this chapter, it has been determinant the scale of work (global) and the availability of homogenous information for entire area of study.

The structure of the system follows a hierarchic scheme with three sub-indices concerning the fire danger associated to ignition, propagation and vulnerability.

In chapters 7 to 9, the variables that would have to integrate each one of the mentioned sub-indices are analysed.

The methods for their obtaining and calculation are proposed.

The input variables include the fuel moisture content (live and dead), the risk of ignition associated to man, models of simulation of fire behaviour, vulnerability related to population, environmental value of territory, and potential soil erosion.

Considering the state of the art and the available information, it has not been possible to analysis all of them with the same detail and precision.

Final conclusions about the integration of the proposed variables and the distribution of the index are included in chapter 10.

Chapter 11 is dedicated to the numerous references and the annex (chapter 12) presents a specific work dealing with mapping the post-fire risk of soil erosion

.

EUFIRELAB

D-08-06.doc 1

1 SCOPE AND OBJECTIVES

The main objective of the Wildland Fire Risk and Hazard Unit of the Eufirelab Project (Unit 8) is the design of the basic structure and characteristics of a prototype of a Euro-Mediterranean Wildland Fire Danger Rating System.

With that intention, the work of this unit has been structured in three main sections: - Revision of the variables used in existing fire danger

indices (deliverables D-08-02 and D-08-07), - Analysis of methods for mapping the wildland fire

danger (deliverable D-08-05), and - Design and development of a prototype of Euro-

Mediterranean Wildland Fire Danger Rating System (deliverables D-08-03 and D-08-06), based on the information analyzed in the previous deliverables.

On the basis of the state of the art (existing systems, necessary inputs, algorithms...) and the information availability on the variables of interest, the consortium is intended to develop a prototype of the Euro-Mediterranean Wildland Fire Danger Rating System, designing the basic structure and determining its characteristics.

This final objective will be developed in two deliverables.

In the first (D-08-03), a proposal of the basic structure of the index and a broad selection of the variables to use is carried out.

In the second one (D-08-06), that we present next, will be define more precisely the structure and the variables to include in an index that can be used at global scale for the Euro-Mediterranean Basin.

According to these objectives, the deliverable D-08-06 "Towards to Euro-Mediterranean Wildland Fire Danger Rating System" is carried out in three phases.

1.1 FIRST PHASE

First includes a review of methods already developed and applied operationally or semi-operationally for estimating and mapping wildland fire risks at Euro-Mediterranean scale (Megafires and Spread projects, European Forest Fire Information System-Risk Forecast - EFFIS -).

In this section it has been realized a detailed analysis of the proposals, at global scale and for the Euro-Mediterranean basin, about which we have been able to find information.

EUFIRELAB

D-08-06.doc 2

1.2 SECOND PHASE

In the second phase is proposed the basic structure of this risk index to integrate the most important components in the ignition, propagation and vulnerability of forest fires.

This was carried out taking into account the proposals and suggestion made in these previous projects and the analysis performed in preceding deliverables of this work unit (8), concerning variables of risk and methods of integration.

In this section it has been determinant the scale of work (global) and the availability of information (more or less homogenous) for entire area of study.

These two premises have supposed the proposal of a relatively simple structure, which could become more complex and rich in local scales of work as well as in the case of having available more homogenous information about variables of interest for whole area.

Even so, as we see in the last section of this deliverable, it is not at all simple to propose an operative index for a region so extensive and diverse (natural and culturally) as Southern Europe.

The structure of this index follows a hierarchic scheme in which the final index is obtained from three sub-indices concerning the fire danger associated to ignition, propagation and vulnerability.

These sub-indices offer independent information that it can be used individually.

For example, the ignition danger index can be very useful for the design of the vigilance and monitoring system, whereas the propagation risk index can be very interesting for the location of extinction resources.

The vulnerability shows us those zones that could suppose greater losses in human, ecological and economic terms.

The combination of these three sub-indices will offer an overall view about the forest fire danger, considering all the key elements that need to be valued in an index of these characteristics.

1.3 THIRD PHASE

In the third phase, after the proposal of this basic structure, the variables that would have to integrate each one of the mentioned sub-indices are analyzed. In some cases the methods for their obtaining and calculation are proposed.

As it has been previously commented, the selected variables will depend significantly on the scale of work and the availability of information.

For this reason, in this section of the deliverable, study cases at local scale are also presented, with the intention of transferring that proposal from the local scale to a global scale.

The input variables in the ignition sub-index include the fuel moisture content (live and dead) and the risk associated to man.

In the sub-index of propagation, models of simulation of fire behavior are included. Finally, variables associated to population, environmental value, and soil erosion, are incorporated in the vulnerability sub-index.

Considering the state of the art and the available information, it has not been possible to take into account all of them with the same detail and precision.

Finally, some recommendations about the integration of the mentioned variables are proposed and some ideas about the distribution of the index are included.

EUFIRELAB

D-08-06.doc 3

2 PROPOSED METHODS FOR WILDLAND FIRE DANGER ESTIMATION AT EUROPEAN SCALE

As it was mentioned in the description of work, the review of the state of the art on fire danger and hazard assessment and mapping is one of the main objectives of Unit 08.

The review was divided in two major parts: - the analysis of the individual environmental and

anthropogenic variables related with forest fire danger and hazard (accomplished in D-08-02), and

- the description of the models and methods currently used to combine those variables to estimate fire danger and hazard and produce maps of their spatial distribution at different scales (completed in D-08-05).

A long tradition of research, both European and worldwide, has been devoted to those issues, specially at local and regional scales, but few attempts has been made in order to create an integrate fire risk assessment system at coarser scales (i.e. Pan-European).

In the above mentioned reviews missing links in the integration and spatial assessment of risk information were observed as most of the works were produced and tested in specific study areas and they didn´t offer much information on how to extrapolate the methods to other areas and, even most important, to other spatial scales.

The objective of this section is to review existing methods, systems or products that have been proposed for fire danger/risk estimation at European scale.

This review will allow us to analyse and compare the structure and components of those systems and also to identify their goodness and, eventually, potential limitations for being operationally applied to evaluate fire risk in the framework of national or supranational decision support systems.

This information is required to cope with the second objective of Unit 08, which is to propose the structure of a Euro-Mediterranean Widland Fire Danger Rating System.

Wildland fire danger may be considered at different spatial and temporal resolutions: global and local scale on one hand, and short-term and long-term on the other.

Both scales of assessments are very important for fire management.

Considering only spatial scale, global approaches assume to cover from millions to dozen million square kilometres, while local studies are focused on hundreds to few thousands square kilometres.

Global analysis can contribute to establish general guidelines for fire management at international level, such as at European Union level, while local scales are adapted to specific fire prevention resources of small regions or provinces.

The type of information to be used in fire risk assessment systems is strongly dependent on the spatial scale, since global approaches could most probably not include some critical variables that are easily available at local or regional scale.

As it was previously mentioned, in this review section we will include only those methods for wildland fire danger estimation that had been proposed and effectively implemented at European scale, that is, methods that have produced maps of fire risk/danger for the whole Europe or, at least, an area covering several European countries.

Systems that were theoretically developed for being applied at global scales but were never implemented or only in specific test areas are not included in this review.

We have found three systems that fulfill the above mentioned requirements, two of them are the result of european projects “Megafires” and “Spread” and the last one is a fire risk forecasting system developed and implemented by the Joint Research Centre (JRC) in the framework of the INFOREST Action (http://inforest.jrc.it/effis/).

EUFIRELAB

D-08-06.doc 4

3 REVIEW OF FIRE DANGER RATING SYSTEM DEVELOPED IN MEGAFIRES PROJECT

3.1 INTRODUCTION

One of the objectives addressed in the MEGAFiReS project (http://www.geogra.uah.es/megafires/) was to develop methodologies for fire risk rating and mapping, at both global and local scales, based on satellite remote sensing and Geographic Information System (GIS) techniques.

For the global scale, it was intended to present a comprehensive view of fire risk assessment at the European Mediterranean countries.

MEGAFiReS project undertook, for the first time in Europe, a cross-country evaluation of trends in fuel moisture content of Mediterranean species using coarse resolution remote sensing data.

Three different aspects of fire risk assessment were investigated in MEGAFIRES: - Short-tem fire risk: foliage moisture estimation from

satellite data - Fire danger rating using meteorological indices and

satellite data - Long-term fire risk mapping

Fuel moisture estimation was investigated in MEGAFIRES using high and low resolution sensors.

This was one of the first attempts to correlate daily-acquired low resolution satellite data with moisture content of living foliage on a significant European scale.

Validation of the satellite-derived information was based on the measurements of foliage moisture content of different species carried out during 2 years on 4 networks of test-land plots located in three different Euro-Mediterranean countries.

No Pan-European maps were produced and, therefore, we will not describe here the proposed methodology.

However, the study clearly showed the potential of low-resolution satellite data such as NOAA-AVHRR to monitor and map foliage moisture content at global scales.

On the contrary, the other two aspects of fire risk that were included in the project, produced methods and results at Euro-Mediterranean scale and, therefore, they will be described below.

3.2 USE OF METEOROLOGICAL INDICES AND SATELLITE DATA

In Europe there is no uniform approach to fire danger rating and different methods have been applied and developed in different Mediterranean countries.

Some studies were made in Minerve I and II European research projects to compare the fire danger rating capabilities of different meteorological danger indices, in the attempt to find a common model to be applied in the European Mediterranean area, but the final results left some uncertainty about the best model to use (BOVIO et al., 1994; VIEGAS et al., 1996).

One of the problems faced by the danger indices is that they are not specifically tailored to assess extreme fire danger conditions.

In MEGAFIReS a first global attempt was made to develop a system for rating, from ground weather measurements, extreme fire conditions in the European Mediterranean basin.

The system is empirical and is made up of a combined set of selected existing fire danger indices.

On the other hand, the possibility of using satellite data to estimate the variables needed to compute the danger indices specifically tailored for large fires was investigated.

The objective was to have a more accurate spatial distribution of the fire danger estimation that could be less dependent on the location and density of the weather stations.

A global historical fire and meteorological databases were addressed for the whole European Mediterranean basin.

Meteorological variables required to compute fire danger indices were obtained from the database of the MARS project implemented at the Joint Research Center.

In this database, in the time period considered, daily ground weather data recorded at 12:00 a.m. from about 360 weather stations were interpolated on 1389 grid cells of 50x50 Km (Figure 1).

Historical meteorological data included maximum and minimum air temperature, vapour pressure, wind speed, rainfall, potential evaporation and solar radiation.

With these weather data, 28 fire danger indices were computed for each grid cell (daily values for five years).

The indices computed had different features, addressing specific components of fire danger (see table 1).

Due to the high heterogeneity of the fire environment in the territory, the area of investigation was restricted and stratified using land cover and climatic data.

EUFIRELAB

D-08-06.doc 5

Only those grid cells with at least 25% of their area covered by wildlands (i.e. the sum of grassland, shrubland and forest land according to the CORINE Land Cover map legend) were retained for further investigation, so those grid cells mostly covered by urban and agricultural lands were masked out.

In addition, only the Köppen climatic types BSh, BSk, Csa and Csb were considered truly representative of the investigated fire environment.

The other climatic types, except for some transitory situations, were in different conditions, often with a winter-spring fire season and a generally lower level of fire danger.

In Figure 2 the area selected for the investigation, stratified in the climatic types is showed.

The restriction of this defined area of interest resulted in a reduction of the grid cells from 1389 to 499, which multiplied by the number of days considered (5 fire seasons of 122 days each) give a total of 304 390 statistical units.

The empirical model was build using logistic regression in order to identify extreme/non extreme fire danger conditions.

The objective was to identify the set of indices that together could be used to best estimate the meteorological conditions for a Danger Day (DD).

It must be pointed out that large fires often last for more than one day, that is to say that meteorological danger conditions associated with them remain for a certain period.

Actually, it often happens that a large fire that burns for several days spread most of the burned area in few highly severe days.

In order to consider a reasonable statistical unit an approximation was introduced defining as Extreme Danger Day (EDD) a day with at least 1 large fire burning in the grid cell, excluding multi-day burning fires and the day when the fire was extinguished, that is the last day, when is was reasonable to expect a lower level of fire danger.

The dependent variable in the logistic regression was a binomial variable that identified an EDD (value 1 as danger day or 0 as non danger day), while the explanatory variables were a selected set of danger indices that provided the best fit.

Exploratory data analysis of indices led to identify and exclude some indices with undesirable distribution properties.

After test and build models, better results were obtained when separating the 3 climatic zones i.e. building different models in each zone and including indices of different types in each model (short, mid and long-term moisture content estimators, spread potential indices and composite indices), showing that when different components of fire danger are taken into account the estimates are improved.

All models were globally significant at the 1% level and so were for individual variable coefficients, with some exception.

The negative signs found in coefficients were for indices that directly indicated moisture content values, while positive signs were for indices that increased with danger level.

In the typical Mediterranean climatic types (Csa and Csb classes) the models showed better performances, while in the semi-arid zone (BS class) predictions were more uncertain.

Nevertheless, in all zones and, therefore, at global level, the logistic models developed showed better performances of each of the 28 individual fire danger indices considered in the study.

The indices finally selected should not be considered as being generally better than the others. Exploratory data analysis showed that many indices that were not included in the equations had good discriminating power.

Those selected were a “set” of danger indices that together improved the rating capabilities of meteorological severe fire conditions.

To assess the model fitting, observed and predicted EDD were compared.

This was done grouping the observations according to different criteria (probability classes, climatic zones and month) and also applying the performance score for fire danger indices introduced by (MANDALLAZ and YE 1997)

The spatial fitting of the model to the data was qualitatively assessed by means of monthly maps of DDs plotted over the monthly values of the model output in each grid cell. One example of these monthly maps is presented in figure 3.

The agreement between the spatial distribution of the estimated fire severity and large fires ocurrence, changed from year to year, but an overall correct behaviour of the logistic model was recognised.

The logistic regression analysis provided a first approximation to underline the meteorological indices better adapted to predict large fire occurrence.

The authors proposed the logistic model as a first prototype that must undergo further analysis.

Besides, the authors consider the usefulness of incorporating other refinements such as the integration of satellite images for improving the spatialization of the variables.

Within MEGAFIRES project, the potential use of satellite data for the spatial extrapolation of specific components of meteorological fire danger was explored.

The indices considered were the Keetch-Byram Drought Index (KBDI; KEETCH and BYRAM, 1968) and the Canadian Drought Code (DC; VAN WAGNER 1987).

The analyses were performed at three different levels, in a region of Spain, in the Iberian Peninsula and in the whole European Mediterranean basin.

Both temporal and spatial dimensions were addressed when analysing correlation of satellite variables with meteorological danger indices.

EUFIRELAB

D-08-06.doc 6

For Andalusia and the Iberian Peninsula the results confirmed the correlation between NDVI and KBDI was quite poor, but the one between NDVI and DC was interesting.

For the global approach both meteorological danger indices (DC and KBDI) were computed with records of the JRC-MARS database for a selection of 56 weather stations.

The satellite data were the same mosaic of NOAA-AVHRR images used from the correlation analysis with the logistic model, with forested area masked out in circles of 10 km radius around selected weather stations.

Regarding the temporal analysis, the expected signs of correlation between NDVI and DC (inverse relation) were obtained for 44 out of 56 weather stations.

Due to the few dates available, the statistical significance at the 5% level could be established for only 9 of them.

Regarding the spatial dimension, the correlation between NDVI and DC calculated for all available dates showed the expected trend for 31 dates out of 35.

The strongest correlation period, from 19 to 23 July, with R=-0,74 and 92 records of available data, was then used to develop a regression equation to derive DC values as a function of NDVI (figure 4).

The common background of some meteorological fire danger indices and remote sensing-derived indices was confirmed by observation at the regional level and partially at the European Mediterranean basin level, and this is particularly important for the long-term components of fire danger.

Furthermore, the DC, better than the KDBI, seemed to be more related to living vegetation vigour and its seasonal trend in Mediterranean environment, as it is monitored by remote sensing techniques.

These consistent similarities, regardless of the local climatic differences, strongly support the use of satellite data to estimate temporal trends in some components of meteorological fire danger, and encourage further investigation on its potential integration into currently operational fire danger indices.

3.3 FIGURES

Figure 1: Layout of grid cells in the European Mediterranean Basin and location of large summer fires 1991-1995

Figure 2: Area of investigation (study area) stratified in the climatic types

EUFIRELAB

D-08-06.doc 7

Figure 3: July 1994 DDs and model values in the study area

Figure 4: Logistic model output estimated from ST in the European Mediterranean basin (July 19-23, 1996).

The mapped area is the one for which the logistic model was developed, i.e. the climatic zones Csa, Csb and BS, with urban and agricultural 50x50 km2 grid cells masked out

EUFIRELAB

D-08-06.doc 8

3.4 TABLES

Table 1: Indices used in the study

Category N Index Description Variable

1 BEHAVE fine fuel moisture model (ROTHERMEL et al. 1986) BEHAVE

2 Canadian Fine Fuel Moisture Code (FFMC) (VAN WAGNER 1987) FFMC

3 Moisture content estimate in Mark5 McArthur’s Fire Danger Meter (NOBLE et al) Mark5

4 McArthur ‘67 fuel moisture model (MCARTHUR 1967) McArth67

Short-term moisture content

estimators

5 USA NFDRS 1-Hour Timelag Fuel Moisture (DEEMING et al. 1978) h1

6 NFDRS 10-Hour Timelag Fuel Moisture (BRADSHAW et al. 1983) h10

7 NFDRS 100-Hour Timelag Fuel Moisture (BRADSHAW et al. 1983) h100

8 Fosberg’s et al. 1000-Hour Timelag Fuel Moisture (FOSBERG et al. 1981) h1000

9 Canadian Build Up Index (BUI) (VAN WAGNER 1987) BUI

10 Canadian Drought Code (DC) (VAN WAGNER 1987) DC

11 Canadian Duff Moisture Code (DMC) (VAN WAGNER 1987) DMC

Mid-term and Long-term moisture content

estimators

12 Keetch-Byram Drought Index (KEETCH and BYRAM 1968)- revision, 1988) KB

13 Canadian Initial Spread Index (ISI) (VAN WAGNER 1987) ISI

14 Mark3 McArthur’s Fire Danger Meter (NOBLE et al. 1980) Mark3

15 Mark5 McArthur’s Fire Danger Meter (NOBLE et al. 1980) Mark5F

Spread potential

estimators 16 Italian Fire Danger Index (PALMIERI et al. 1992) Impi

17 Canadian Fire Weather Index (FWI) (VAN WAGNER 1987) FWI

18 Portuguese Index (GONÇALVES and LOURENÇO 1990) PortoCif

19 Spanish ICONA Method – probability of ignition (ICONA 1993) Prob

20 Orieux Index (ORIEUX 1979) Orieux

21 Carrega ’87 Index (CARREGA 1990) Carrega

22 Sol Numerical Risk (SOL 1990) SolRisNum

Composite indices

23 IREPI index (BOVIO et al. 1994) IREPI

24 Anderson et al. adsorption (ANDERSON et al. 1978) EWmax

25 Anderson et al. desorption (ANDERSON et al. 1978) EDmax

26 Van Wagner adsorption (VAN WAGNER 1982) EWmaxW

27 Van Wagner desorption (VAN WAGNER 1982) EDmaxW

Equilibrium Moisture Content

(EMC) models

28 Simard’s Equilibrium Moisture Content equations (SIMARD 1968) SEMCmax

EUFIRELAB

D-08-06.doc 9

3.5 ANALYSIS OF LONG-TERM FIRE RISK ON A EUROPEAN LEVEL

The second MEGAFIRES task regarding risk assessment at Euro-Mediterranen scale was related with the definition of an integrated long-term fire risk index.

The main goal of this analysis (CHUVIECO et al. 1999b) was to identify structural factors of fire risk at European scale.

As in the case of meteorological indices, fire occurrence was focused on large events (above 500 hectares).

The basis for the analysis was the compilation of a database for the whole study area (Greece, Italy, South of France, Spain and Portugal).

The variables were obtained from National or Global databases: CORINE, Digital Chart of World, NGDC Globe project, DMSP data.

Census data were the most difficult to compile, due to the problem of ensuring consistency among the different countries.

Census data defined the geographical unit of reference.

Since most of the human variables were only available at a provincial level (NUT-3), this division was used for all the variables.

In the case of geographical layers which cover the whole territory (such as land cover, elevation and DMSP data), the average value for each province was computed to assure consistency among variables.

Within the study area, there are 164 NUT-3 provinces: 48 in Spain, 20 in Italy, 52 in Greece, 18 in Portugal and 26 in France.

Provincial average sizes vary from one country to another, Italy having the largest and Greece the smallest.

3.5.1 Selection of risk variables

The selection of variables relating to long-term trends of fire occurrence was carried out using previous technical literature reviews, (CHUVIECO et al. 1997) and considering the limitations of data among the different countries.

Three groups of variables were identified: first the geographical ones, that deal with terrain features (climate, land cover, roads, rails, etc.); secondly the demographic variables, which were related to population characteristics, and finally agricultural variables, dealing with agricultural structure.

Some variables that could be, at least theoretically, related to human risk, such as property size, unemployment, and hunting practices were not considered, because of the difficulty in either obtaining data or homogenising them among the different countries.

Since rural economies have strongly changed in most Mediterranean countries during the last 30 years, dynamic variables were also included in the analysis, by comparing present values with those measured in 1960.

A whole set of 52 variables taken from demographic and agricultural census were extracted for 164 provinces of the study area.

Regarding the geographical variables, a brief description of the spatial analysis undertaken follows:

3.5.1.1 Elevation

Elevation was obtained from two sources: National maps at a 1:1,000,000 scale for Spain and Italy, and the GLOBE project (compiled by the U.S. Geological Survey at 1 Km2 resolution). From the elevation data (Figure 5), mean slope and roughness were computed using algorithms provided by Idrisi G.I.S. (EASTMAN 1993)

3.5.1.2 Land cover

Land cover, which is one of the layers generated for the European Environmental Agency (European Environtmental Agency 1996) was extracted from the CORINE program.

Since some regions were missing from the published CD-ROM, the coverage was completed with data directly provided by the EEA.

The European legend of the CORINE land-cover program was simplified to six general fuel type categories: Grasslands, Shrublands, Perennial, Broadleaf, Agriculture, and Non-Vegetated (Figure 6).

3.5.1.3 Roads and railways density

The density of roads and railways was computed from national maps and the Digital chart of the World.

3.5.1.4 Urban areas

Urban areas were assumed to be related to fire risk, since recreational uses of forests are among the main causes of fire ignition, which can be either accidental or due to carelessness.

Since urban areas are very dynamic, satellite information was chosen as a source to map urban land cover.

Data from the Defence Meteorological Satellite Program (DMSP) provide a global view of city lights, since this satellite includes a very sensitive radiometer operating at night in the visible spectrum (Figure 7).

The National Geophysical Data Center (NGDC) in Boulder, Colorado, has processed these data to generate a world database of city lights. (ELVIDGE et al. 1997).

City lights data from Eurasia were extracted from the Internet server at NDGC.

3.5.1.5 Climatic regions

Climatic regions were generated from the Joint Research Center climatic database archived at the MARS unit.

Average values for the last 30 years were used to classify each cell, according with the Koeppen method.

The original Koeppen system was reduced to a smaller number of classes and those not presented in the MEGAFiReS study area were discarded.

EUFIRELAB

D-08-06.doc 10

The final map includes the following climates: BSh+Csa, BSk, Cfa, Cfb+Cfc, Csb+Csc, Dfb+Dfc.

The proportion of each provincial unit per climate type was extracted after a cross-tabulation between the Koeppen system and the provincial limit.

3.5.2 Techniques to estimate large fire occurrence

The number of large fires and burned area in each province of the EU Mediterranean countries was compiled from national statistics (Figure 8).

The study period covered was from 1991 to 1995.

Presence of large fires in each provincial unit was selected as the target variable used to estimate fire occurrence from the auxiliary variables previously mentioned.

The final objective was to better understand how these variables explain patterns of large fire occurrence in the EU Mediterranean basin.

Two techniques were selected to perform this analysis: Logistic regression and Artificial Neural Networks.

Both techniques had been previously used in the estimation of fire occurrence providing coherent results.

Additional analysis was performed with the number of large fires per provincial unit.

In this case, linear regression analysis was used instead of logistic regression, since this latter technique can only be applied to dicotonomous variables (fire/not fire).

Neural networks were applied both to the presence or absence of fires and the number of fires per provincial unit.

3.5.2.1 Estimation using logistic regression

Different models were tested including different set of variables.

In order to avoid the effect of multiple correlation and the noise produced by the large number of variables, a previous analysis was carried out.

Multiple correlations were computed after grouping the variables into three categories (geographical, demographic and agricultural variables).

These variables were input into three different LR, one for each of the above groups, in order to find out the most significant variables to explain fire occurrence.

In all cases, the occurrence of large fires (0 not affected, 1 affected) was used as dependent variable.

LR was performed with the Stepwise Backward Selection algorithm included in the SPSS statistical package.

In order to build a global LR model, which includes the geographical, demographic and agricultural variables, a new correlation matrix was computed among them, discarding some variables.

The final model was computed from 161 cases. Three were discarded because they offered a high

bias.

The variables finally selected were: LDENSI90 (population density), POBACT90 (Active population 90), LALTIMED (mean elevation), LDECIDU (% of broadleaf forest), and CLIM_BCS (Proportion of area with Koeppen climates B and CS).

Table 2 provides the LR coefficients for the final variables and their level of significance.

The signs obtained were logical since the fires are expected to be higher at greater population densities, higher elevations, more arid climatic conditions, less active population and less area covered by broadleaf trees.

The variables considered in the equation offer trends in the correlation with the dependent variable according to what was expected (see column in table 2).

The significance level varies among them, being more important for active population, climate and population density, which should be considered as the most related to large fire occurrence

Table 3 offers an assessment on the performance of the model.

A global accuracy of 78.26% was obtained in the estimation of large fires between 1991 and 1995.

This prediction is quite acceptable, considering the great diversity of the study area, not only in a geographical sense but also taking into account the different national particularities regarding fire ocurrence.

This performance is even better if only omission errors are considered, since only 11% of the provinces where fires occurred were not classified as such.

Commission errors were higher (almost 40% of the provinces predicted as having a large fire were not affected), but these errors are less critical than the omission’s from a fire management point of view.

The LR function derived from the final model makes it possible to compare the geographical distribution of expected versus observed occurrence of large fires (figure 9).

Most provinces were predicted correctly.

To perform a final test on the LR model, a new equation was generated from a random sample of 60% of all the provinces.

The other 40% of provinces were used as test cases.

The LR model was created following a similar approach to the previous one.

A threshold below 0.5 was also selected to discriminate between occurrence / not occurrence.

In this case, as could be anticipated, the fitting is poorer that in the previous equation (table 4).

A global accuracy of 60% was achieved, which could be considered a good estimation, especially taking into account the rate of omission errors (37,5%).

The same variables as the previous model were identified as significant to explain fire occurrence in these test provinces.

Geographical distribution of the estimations is included in Figure 10.

EUFIRELAB

D-08-06.doc 11

3.5.2.2 Estimation using linear regression

Linear regression analysis was applied to estimate the number of large fires in each provincial unit.

Following similar criteria to the logistic regression, a previous selection of significant variables in each thematic group (geographical, demographic and agricultural) was undertaken.

Afterwards, exploratory analysis to the resultant variables was carried out to assure normal distribution of the independent variables.

Some were log-transformed. Criteria for selecting the independent variables were

based on iterative stepwise forward method with thresholds of significance at 0.05.

The final model was:

NF= 13.467 - 0.273 POBACT90 + 0.161 DIFPOBACT + 0.051 BSHCSA - 0.113 DENCAP90 + 0.048 DENOVI90 - 0.060 AGRICULT -0.073 DIFREN,

where - NF is the number of large fires; - POBACT90 is the percentage of active population in

each province for 1990; - DIFPOBACT is the difference in active population

between 1960 and 1990; - BSHCSA is the proportion of provincial area

covered by KÖPPEN climates Bsh and Csa; - DENCAP90 is the density of goats for 1990; - DENOVI90 is the density of sheeps for 1990; - AGRICULT is the proportion of agricultural cover in

each province and - DIFREN is the difference in agricultural renters

between 1960 and 1990. The RMS error of this estimation was 5.4 fires.

3.5.2.3 Estimation using artificial neural networks

The predictive power of artificial neural networks (ANN) was compared with logistic regression to estimate large fires ocurrence (CHUVIECO et al. 1998 and 1999, CARVACHO 1998).

The independent variables were the same (human factors, climatic units, vegetation classes and topography) used in the regression models.

Also, the same provinces randomly selected to fit the logistic regression were used to train the ANN.

Then, the trained model was applied to the other 40% of the provinces, studying the predictive capacity of this technique in comparision with logistic regression.

Figure 10 includes the spatial assessment of both predictions., where is displayed the estimated versus observed provinces with fires/no fires.

Table 5 includes the error matrix of ANN analyses.

In the final map (figure 10), red and yellow colors indicate correct predictions (either estimated or observed or non estimated and non observed), while orange and magenta the errors.

The most critical are omission errors (displayed in magenta) where the models did not predict fires in provinces that were actually affected

After a comparative analysis, the omission errors (the most important ones) were higher for the logistic regression, having also lower global accuracy than ANN.

The network work better for typical Mediterranean fires (those occurring at summer time), while it failed in those provinces affected by winter fires (provinces located in the North of Spain, Greece and Italy, and some in South of France).

Errors from logistic regression did not present a clear spatial pattern, affecting provinces of different fire characteristics.

In spite of the flexibility and fitting power of the ANN one of the main drawbacks of this technique refers to complexity to find out what are the most significant variables that affect fire occurrence.

In this aspect, the ANN is like a “black box” with few analytical possibilities to measure the influence of the independent variables in the estimation.

An indirect method to find out the most critical variables of the model was undertaken in the Megafires project.

This method is based on replacing the original values of each input variable by random values, after the network is trained.

It was assumed that the increase in the RMS error produced by such a change should reflect the relative importance of that variable in the whole fitting.

If the variable was significant, randomising it should give us a greater RMS with respect to a marginal one.

Repeating this step with each variable showed the relative importance of all of them.

Randomising Csb-Csc climates, average distance to the roads, active population in 1990, agricultural area, differences in population working in the services between 1960 and 1990 and population density provided the highest increment in RMS.

Therefore, these should be the most critical variables in the estimation of fire-not fire.

3.5.2.4 Conclusions

According to the Logistic Regression model the variables most clearly related to large fire occurrences was: the proportion of BS-Cs climates, population density, elevation, unemployment and lack of broadleaf cover.

The ANN offered a robust estimation of fire occurrence, but didn't provide as many insights on the most critical variables.

Estimated fire occurrence maps proved to be a useful tool for managing global trends of fire risk, by pointing out those regions, which offer a more consistent and stable risk of being affected by large fire events.

Evident difficulties found were related witht the availability and homogeneity of socio-economic data at European Scale.

EUFIRELAB

D-08-06.doc 12

3.5.3 Figures

Figure 5: Elevation of theMegafires study area

Figure 6: Fuel type map generated from the Corine Land Cover database

Figure 7: DMSP city lights map of the study area (Source NGDC

EUFIRELAB

D-08-06.doc 13

Figure 8: Occurrence of large fires (above 500 hectares) in the study area 1991-1995).

Figure 9: Actual versus predicted occurrence of large fires by LR techniques 1991-1995)

Figure 10: Observed versus estimated occurrences of large fires in Southern Europe using logistic regression

analysis (top) and neural network analysis (below).

Est-Obs means fires that were both estimated and observed; Nest-Nobs, means both non-estimated and non-observed; Est-Nobs and Nest-Obs imply commission and omission errors, respectively

EUFIRELAB

D-08-06.doc 14

3.5.4 Tables

Table 2: Coefficients of the final logistic regression model. Dependent variable occurrence/Not occurrence of large forest fires 1991-1995)

Variable B S.E. Wald df Significance R Exp(B)

LDENSI91 .8349 .2564 10.6024 1 .0011 .1963 2.3045

POBACT90 -.2168 .0500 18.7953 1 .0000 -.2743 .8051

LALTIMED .6631 .2245 8.7219 1 .0031 .1735 1.9408

LDECIDU -.3524 .1572 5.0240 1 .0250 -.1164 .7030

CLIM_BS .0229 .0063 13.1270 1 .0003 .2233 1.0232

Table 3: Observed versus predicted cases

Predicted (cases) % Correct

Observed (cases) 0 1

0 38 25 60.32 %

1 10 88 89.80 %

Overall 78.26 %

Table 4: Error matrix for the estimation of large fire occurrence using logistic regression (observed versus predicted cases. Test provinces)

Observed

Predicted No Fire Fire Totals Comission error

No Fire 14 15 29 51,72%

Fire 11 25 36 30,56%

Totals 25 40 65

Comission error 44 % 37,5% Global accuracy 60,23%

Table 5: Error matrix for the estimation of large fire occurrence using ANN Observed. Test provinces

Observed

Predicted No Fire Fire Totals Comission error

No Fire 17 12 29 41,39%

Fire 8 28 36 22,22%

Totals 25 40 65

Comission error 32% 30 % Global accuracy 69,23%

EUFIRELAB

D-08-06.doc 15

4 REVIEW OF FIRE RISK ESTIMATION METHOD DEVELOPED FOR THE SPREAD PROJECT

The Spread project (http://www.adai.pt/spread/) is a European funded project which tried to assess fire risk conditions at several spatial scales, including the Pan- European.

Following Spread Deliverable D162 entitled “Fire risk mapping (II): Validated methods and digital products covering the whole EU Mediterranean Basin and selected study sites (local – regional – global scale)” we present here a summary of the fire risk integration scheme proposed for the European Mediterranean area (EUMed).

Acording to previous literature reviews on fire risk terminology, the Spread project studied the fire risk in terms of physical probability of fire occurrence (danger), on one hand, and potential effects caused by the fire on the other (vulnerability) (CHUVIECO et al. 2003b).

The former refers to the potential that a fire occurs in a particular area and time on one hand, and to its propagation capability on the other.

The other component of fire risk is named fire vulnerability, and concerns the potential effects of a fire, either on human values and lives and environmental resources.

The final index should be computed as the product of the two components

Trying to adopt an approach that may be both scientifically sound and operationally applicable, the components that include those two concepts of danger and vulnerability were adapted within Spread Project to different spatial and temporal scales, but it was more focused on the EUMed scale.

The danger component was considered in a broad perspective, covering the probability of a fuel ignites (ignition danger) and the potential hazard that this fire propagates in space and time (propagation danger).

The consideration of both components forms the wildland fire danger assessment (WFDA) component of the WFR index defined by the Spread project, as illustrated in Figure 11.

Therefore, the WFDA is based on the estimation of two properties: ignition danger and propagation danger.

The former is related to the causal agents of fire as well as the conditions and properties of the fuel, while the latter is associated to estimating the behaviour of the fire.

Fire behaviour is mainly described with reference to the propagation rate and the intensity of the flame front.

The fire danger assessment system includes an estimation of fuel moisture content derived from satellite data (live fuels) and meteorological variables (dead fuels), as well as an estimation of the historical patterns of human-caused fires, and the fire propagation potential, generated from fire-behaviour simulation programs.

Other factors, such as risk associated to lighting or flammability, as well as the vulnerability component could not be derived for the whole EUMed, and was not addressed in this project.

All the input variables were geographically referenced and included into a dedicated Geographic Information System (GIS).

The results were derived for the whole EUMed area at 1x1 km grid size (with the exception of meteorological data that were only available at 50x50 km2 grid sizes, and NOAA-AVHRR images with 4.4 x 4.4 km2 pixel size).

The resulting product showed promising potential for helping fire managers to simulate different danger scenarios, as well as to obtain a single evaluation of fire danger conditions for the whole EUMed area size.

A demo was run as a Spread web service during summer 2004, in which only the fire danger branch (WFDA) of the framework was implemented and dynamically updated in quasi-real time during 1 month.

EUFIRELAB

D-08-06.doc 16

4.1 INPUT DATA FOR FIRE RISK MAPPING

Once provided a general explanation of the scheme adopted in the Spread project, a more detailed definition of each variable and the data inputs for generating the index at EUMed scale follows in this section.

The WFDA is a combination of the following intermediate maps:

- HDI: Human Danger Ignition (static) - PD: Propagation Danger (static) - PI: Probability of Ignition related to fuel moisture

content (dynamic))

Static means that the intermediate map does not change throughout the season, dynamic means that the intermediate map will have to be updated every day. So the final Wildland Fire Danger Assessment map (WFDA) has to be updated daily.

4.2 HUMAN DANGER IGNITION

The human danger ignition HDI layer was generated from historical data of fire occurrence.

A probability surface was generated from the location of ignition points using the kernel density approach.

Kernel density estimation is based on the estimation of the density at each intersection of a grid superimposed on the data, after placing a probability density (kernel) over each point event (GATRELL et al. 1996, LEVINE 2002).

KOUTSIAS et al. (2002, 2004) introduced it in fire occurrence for assessing fire occurrence patterns at landscape level by addressing some of the inherent positional inaccuracies of the fire ignition locations.

The number of fires observed at community level for the period 1992-2000 was calculated from national fire statistics and expressed using the community centroids.

The adaptive kernel density estimation mode was chosen due to the non-homogeneous spatial distribution of community centroids.

The adaptive approach allows for the adjustment of the bandwidth size in relation to the concentration of the interpolated points (WORTON 1989).

Locally varying bandwidth size of 10 community centroids proved to perform best showing a reasonable variability in the resulting density surfaces avoiding an under- or over-smoothing.

To avoid over- or under-estimation within and among countries because of heterogeneities of different sources the kernel density values within each country were processed before merging them.

Actually, kernel densities were reclassified to 10 classes based on the equal area criterion within each country, presupposing equivalence for fire hot spot areas among the countries

The kernel density interpolation produced continuous, fire occurrence density surfaces, which in the Mediterranean context is mainly related to human fire danger factors, and therefore can be considered as a static representation of fire danger associated to human factors thorough a whole fire season.

Figure 12 shows the results of this analysis, which provides a global view of fire ocurrence distribution, which is mainly caused by human factors, with higher occurrences in the NW of Spain and Portugal, SW of Italy and Greece, the Southern part of the Maritime Alps, and most territory of Corsica and Sardinia

EUFIRELAB

D-08-06.doc 17

4.3 PROBABILITY OF IGNITION AND FUEL MOISTURE CONTENT

4.3.1 Fuel Moisture Content of live fuels

The estimation of Fuel moisture content (FMC) of live fuels was based on an empirical formula derived from satellite data.

The formula was generated for grasslands and shrublands based on the analysis of multitemporal series of NOAA-AVHRR images. (CHUVIECO et al. 2004b)

The index was derived and tested in Central Spain. For the calibration study, AVHRR images were acquired daily at the University of Alcala’s HRPT receiving station, but 8-day composites were created to avoid cloud and atmospheric contamination.

The compositing criterion was maximizing surface temperature of the daily series.

The empirical index was based in two different equations, one for grasslands and the other one for shrublands (Cistus ladanifer was the species selected for the calibration phase).

For extending this experience to the whole area covered by EU Mediterranean countries, it was used the 10-day NDVI composites of MARS NOAA-AVHRR images produced by the Joint Research Centre (JRC).

The composites were derived from LAC data resampled from the original 1.1x1.1 nadir resolution to 4.4x4.4 km2 pixel size.

To maintain the original time sequence used for developing the empirical formula for FMC estimation, in the application to the EUMed area the 10-day composite was updated every 8 days.

In addition auxiliary information on the extents of grasslands and shrublands in each pixel needed to be derived.

This information was extracted from the Corine land cover map of Europe.

Figure 13 shows an example of the FMC generated from AVHRR images for the whole EUMed area.

4.3.2 Fuel Moisture Content of dead fuels

This map is based on the 10h moisture code of the USA National Fire Danger Rating System, which requires a basic set of input meteorological variables (air temperature and relative humidity).

It was shown that this index could be used to estimate the average FMC of dead fuels in Central Spain (AGUADO et al., in preparation), the area used as calibration site.

Additional testing should be done in other Mediterranean regions for operational use.

The index was computed every day with weather observations taken at noon.

Moisture content was expressed as percent of dry weight.

The meteorological data to compute the index were extracted from the MARS database maintained at the JRC.

The spatial resolution of this database is 50x50 km. The index was computed daily and transferred to

the Internet mapserver designed for the semi-operational testing of the summer of 2004.

Figure 14 shows the spatial distribution of FMC estimated from meteorological variables for the same period as for the figure 13 (in this case, just a single day, instead of an 8-day period).

As it can be observed, the spatial patterns are similar in the two figures, in spite of being generated from two very different sources

4.3.3 Probability of Ignition related to the FMC

This PI index was derived from a combination of the two previous products.

Both dead and live FMC were converted first to probability of ignition, based on average values of moisture of extinction (ME). (CHUVIECO et al. 2004a).

The values of ME are dependent on the fuel complex, and therefore a fuel type map with a ME parameter defined for each fuel type is required to derive this index.

Since no fuel map of Europe with such a specification is yet available, an estimation of fuel type distribution was extracted from the CORINE land cover and a ME value assigned to each fuel type.

Once the FMC of dead and live fuels and their respective PI were obtained, a global PI for each pixel was computed following:

PIf = PIlive * Live proportion + PIdeath * Death proportion

Live and Dead proportion expresses the fraction of live and dead fuel particles in each pixel, which is also dependent on the fuel type.

As a first approach, this map was again generated from the CORINE land cover map.

Figure 15 shows an example of the PIf map computed for the same days as shown in previous figures.

It should be considered as an integration of the fuel moisture status of dead and live fuels.

The spatial resolution of the AVHRR data improves the resolution provided by the MARS meteorological database, and therefore, geographical patterns are more evident in the final product.

EUFIRELAB

D-08-06.doc 18

4.4 PROPAGATION DANGER

4.4.1 Average Rate of Spread and Flame Length

In order to obtain a global view of risk associated to fuel loads, terrain characteristics and wind flows, a global simulation analysis was performed.

This analysis tried to obtain average values of rate of spread and flame length, considering different wind and topographic conditions for the estimated fuel maps of the whole EUMed area.

This attempt should be considered as a general overview of average expected fire behaviour at global scale, in order to rank different danger levels according to the combination of fuel and terrain spatial patterns.

The estimation of the average RoS was based on several simulations performed by the Autonomous University of Barcelona for different fuel types, slope ranges and wind flows.

As a simulation kernel the wildland simulator Bevins proposed by COLLINS, which is based on the fireLib library (COLLINS, 1996) was used.

FireLib is a library that encapsulates the BEHAVE fire behaviour algorithm (MORGAN et al., 2001).

In particular, this simulator uses a cell automata approach to evaluate fire spread.

The terrain is divided into square cells and a neighbourhood relationship is used to evaluate whether a cell will be burnt and at what time the fire will reach the burnt cells.

As inputs, this simulator accepts maps of the terrain, vegetation characteristics, wind and the initial ignition map.

The output generated by the simulator consists of two maps of the terrain.

In the first one, each cell is labelled with its ignition time; in the second one, each cell is labelled with its flame lenght.

This information must be used to calculate the rate of spread and an average from among all flame lenght.

To calculate the rate of spread, the distance between the ignition point and each particular cell in the terrain is divided by the ignition time of that particular cell.

This calculation is repeated for each cell in the terrain to determine the maximum value of the rate of spread.

This maximum value is used as the rate of spread for that particular situation.

To provide the propagation danger map, a set of prototype plots was created, considering all the fuel models from ROTHERMEL classification and a certain slope percentage (from 0 to 100%, with a step of 5%).

The total number of plots was 273. Each plot consists of a grid of cells with 11 columns

x 11 rows (each cell measured 328.083 x 328.083 feet).

The ignition point was located in the middle of the plot.

For each plot, many input parameter combinations were used to simulate the wildland fire behaviour and the average rate of spread and flame lenght were also calculated.

The parameters considered for variation were: 1-hr dead fuel moisture, 10-hr dead fuel moisture, 100-hr dead fuel moisture, live herbaceous moisture and wind speed and direction.

Average values of rate of spread (RoS, in m/min) were computed for the different fuel types.

Finally, the RoS values were scaled into a 0-1 range, by normalizing the values between the maximum and minimum values (figure 16).

Fuel types were derived from the Corine land cover. The resulting map is considered static, since no

specific conditions are simulated (wind or FMC), but only general patterns of propagation rates.

Same process as above was performed to compute the average flame length (FL), measured in metres and normalized into a 0-1 scale (figure 17).

4.4.2 Propagation danger versus RoS and FL

It derives from the combination of the two intermediate products previously described: RoS and FL.

The results of the simulations were mapped at EUmed scale using CORINE land cover (reclassified into fuel models) and slope maps

The maps of FL and RoS were then normalized using linear fitting and multiplied to produce PD:

PD = [(RoSi – RoSmin)/(RoSmax - RoSmin)+0.001] * [(FLi – FLmin)/(FLmax - FLmin)+0.001]

A small constant (0.001) was added to avoid zero multiplication in case of minimum values. RoS and FL were considered in this formula of equal importance, although this could be tuned up in future improvements, according to further experience or suggestions.

This map is taken as static, i.e. it will not change throughout the fire season.

EUFIRELAB

D-08-06.doc 19

4.5 WILDLAND FIRE DANGER ASSESSMENT

The final index of Wildland Fire Danger Assessment WFDA proposed in SPREAD integrates the three components previously described, the human danger index (HDI), the probability of ignition associated to fuel Status (PIf) and the Propagation Danger (PD), as follows:

WFDA = (HDI+c) * (PIf+c) * (PD+c)

- HDI and PD are static, they do not change throughout the season,

- c is a constant to avoid multiplying by zeros (0.001 is used),

- PIf changes every 8 days with respect to live fuels and every day with respect to dead fuels.

Therefore the resulting map of WFDA has to be updated daily. An example is shown in Figure 18, derived using the maps depicted in Figures 12, 15, 16 and 17 for August 10th 2004

All the EUMed maps mentioned so far were made available as a demo product in quasi-real time (with 2-3 days delay) within the demo of the Spread Project web based services.

4.6 FIGURES

Wildland FireDangerAssessment

PropagationDanger

Fuel condition

Ignition sources

Moisture content

Natural: LightningIgnition Danger

HumanIntentional

Unintentional

Flammability

Dead

Live

Wind flows

Topography

Fuel properties

Figure 11: Structure and components of the Wildland Fire Danger assessment (WFDA).

Red boxes mean components that have not been finally implemented in the project

Figure 12: Fire ignition danger in Southern Europe: an estimation of human fire danger factors spatial distribution

EUFIRELAB

D-08-06.doc 20

Figure 13: FMC of live fuels for the first week of August, 2004

Figure 14: Estimation of FMC for dead fuels. Map from 10th August 2004

Figure 15: Probability of ignition from FMC of 10th August 2004

EUFIRELAB

D-08-06.doc 21

Figure 16: Estimated average Rate of Spread (normalized values from 0 to 1)

Figure 17: Estimated average Flame Length (Normalized values from 0 to 1)

Figure 18: Wildland Fire Danger Assessment for August 10th 2004.

EUFIRELAB

D-08-06.doc 22

5 REVIEW OF THE EUROPEAN FOREST FIRE INFORMATION SYSTEM-RISK FORECAST

5.1 INTRODUCTION

(EFFIS-Risk Forecast) developed BY the JRC (Joint Research Centre) of the European Commision

The European Commission DG Joint Research Centre set up since 1999 a research group to work specifically on the development and implementation of advanced methods for the evaluation of forest fire risk and mapping of burnt areas at the European scale.

These activities led to the development of the European Forest Fire Information System (EFFIS).

Since the year 2003 EFFIS is part of the Regulation (EC) No 2152/2003 (Forest Focus) of the European Council and Parliament on monitoring of forests and environmental interactions.

All the EFFIS activities are coordinated with DG Environment to reach the final users, Civil Protection and Forest Services, in the Member States (http://inforest.jrc.it/effis/).

EFFIS is aimed to provide relevant information for the protection of forests against fire in Europe addressing both pre-fire and post-fire conditions.

On the pre-fire phase, EFFIS is focused both on the development of systems to provide forest fire risk forecast based on existing fire risk indices, and on the development of new integrated forest fire risk indicators (EFFIS - Risk Forecast).

These indices permit the harmonized assessment of forest fire risk at the European scale.

They may be used as tools for the assessment of risk situations in cases in which international cooperation in the field of civil protection is needed.

Currently, the dynamic forest fire risk forecast indices are available on the EFFIS web site and sent to the Member States Services daily from the 1st of May until the 31st of October.

The fire risk module of EFFIS runs on operational basis 6 meteorological danger indices (Portuguese, Spanish, Sol Numerical Risk, Italian, Canadian FWI and Behave fine fuel moisture).

In addition a Fire Potential Index (FPI) is computed, by integrating: - forecast meteorological data (to estimate fuel

moisture content), - satellite data (to estimate the relative fraction of live

fuels from Relative Greenness) and - a fuel map (to estimate the fuel properties).

In this case maps with a spatial resolution of 4.4 km are generated.

Daily EU maps are produced from May 1st to October 31st (from 2006 it will be from February 1st) processing 0-24 hours and 48-72 hours weather forecast data.

Together they make the core of the currently named EFFRFS.

Expected forest fire risk level is mapped in 5 classes (very low, low, medium, high, and very high) providing 1 to 3 days risk forecasts over Europe with an average spatial resolution of about 40 km.

Once the indices are computed, they are distributed to the civil protection and forest fire services via Internet.

For all indices, maps of past days or averages for a given historical period can also be generated, specifying either a date or a time interval from 1st of May 1st to October 31st (http://inforest.jrc.it/effis/viewer/viewer.html).

The long term as well as the advanced integrated approach that will complet the risk assessment system are currently under further development and are, at the moment, for internal use.

They will be described here as far as the methods, even at an exploratory research phase, have been proposed and developed at European scale in several publications.

EUFIRELAB

D-08-06.doc 23

5.2 METEOROLOGICAL DANGER INDICES

Dynamic indices are short-term indices that assess the probability of fire ignition and spread.

These factors may be derived directly from meteorological variables, or indirectly by the effect that these variables have on vegetation.

Indices that are computed from meteorological variables are referred to as meteorological fire risk indices.

On the other hand, the indices that evaluate the status of the vegetation are the so-called vegetation stress fire risk indices.

As already mentioned, there is not a consensus on which meteorological fire risk performs best for the whole Mediterranean region, although several studies show that the Canadian Fire Weather Index is well suited for the estimation of fire risk for the region (SOL, 1999).

Accordingly, several indices operationally used in southern Europe are computed in the EFFRFS system.

The algorithms in the EFFRFS are those of a system referred to as EUDIC, (BOVIO and CAMIA 2000) although the actual computation of the fire risk indices is now implemented in a distributed geo-database that enables the on-line integration of all the model inputs.

The indices were initially computed from data collected from a network of meteorological stations covering Europe in which data are collected at the station level, and further interpolated to a 50 km by 50 km grid.

At present, these indices are computed from forecast data, and fire risk indices are produced as forecasts for one, two and three days.

Six indices are being computed: - Behave (Rothermel et al. 1986, VAN WAGNER 1987)

based on estimation of the moisture content of fine dead fuel.

- Canadian Fire Weather Index (FWI) (VAN WAGNER 1987)made up of six normalized indices that indicate the daily variation of fuel water content, initial rate of propagation, and quantity of fuel and expected intensity of the flame front.

- Portuguese index (GONÇALVES and LOURENÇO 1990) assesses atmospheric conditions in the proximity of the fuel layer.

- Spanish ICONA method measures probability of ignition (ICONA 1993)based on litter and fine dead fuels moisture content.

- Sol’s Numerical Risk (SOL 1990, DROUET and SOL 1993) related to ignition and propagation.

- Italian Fire Danger Index (PALMIERI et al. 1992) derived from MC ARTHUR's model.

European statistics on forest fire data are used for the calibration and validation of these indices.

All the indices are computed for southern Europe; however individual civil protection agencies will be able to download the indices for its own territory.

5.3 FPI MODEL1.

5.3.1 General presentation

It is a risk index based on the Fire Potential Index created with BURGAN et al 1998), and integrated through the combination of diverse types of static and dynamic variables from different data sources like meteorological stations, satellite images..

This index has a double structural and dynamic component.

The structural component constitutes the information about fuel properties, which is usually used by indices at local and regional scale, but not at global scale, due to the difficulty to elaborate cartography of fuel types at this scale.

The dynamic perspective is offered by the meteorological variables and the remote sensing data, that serve to characterize the hydrological state of the dead fuel and the alive fuel presence.

The main assumption behind the FPI model is that “fire potential” can be assessed if the moisture content of live and dead vegetation is reasonably represented (BURGAN et al 1998).

Therefore, two are the key factors in this model as can be seen in Figure 19:

1) live and dead fuel loadings, 2) dead fuel moisture content..

In the original FPI model, fuel loadings are obtained from a fuel model map, while dead fuel moisture content is estimated using meteorological data.

Satellite data are used to discriminate between living and “cured” fuel at the pixel level.

Thus, the model requires the following data inputs: - A fuel map to define the dead and live fuel loads,

and the extinction moisture values for each fuel type.

- Maximum Value Composites (MVC) of NDVI (Normalized Difference Vegetation Index) to calculate the Relative Greenness (RG); the purpose of the RG in the model is to provide seasonal adjustment in the proportion of the fuel that is live. In other words, in the FPI model, the proportion of live load varies by pixel, as a function of the assigned fuel model, and of the relative greenness. (BURGAN and HARTFORD 1993)

- Meteorological data to estimate the moisture content of the small dead and cured fuels.

Applying the original FPI methodology was conceptually possible, but neither the input formats nor the spatial resolution of the data were comparable to those used in the US.

Consequently, several modifications of the model were tested.

The main differences between the original FPI and its European counterpart concern the input data and the way these data were processed (e.g. formats, interpolation method, and spatial resolutions).

1 Based on Sebastián et al (2002).

EUFIRELAB

D-08-06.doc 24

Another important difference is connected to the use of the fuel map, which in the original model is meant to provide a live-ratio and an extinction moisture value for every pixel.

To run the model, a fuel map had to be created, since such a map did not exist at a European scale.

5.3.2 Input data

The temporal resolution of the FPI index is one day. All the input data have this temporal resolution, with

the exception of the RG, which is computed with 10 days Maximum Value Composites (MVC) of NDVI daily images.

To avoid the uncertainty of pixel location, a 4.4 km pixel window was used when deriving the MVC from which the NDVI is computed (NOAA AVHRR original spatial resolution is 1.1 km).

Other inputs, in particular, the CORINE Land Cover database (CLC), in its raster version, have a higher spatial resolution (0.1 km).

All the interpolated meteorological data have a lower spatial resolution (50x50 km).

It was decided that the resolution of the NDVI data would be used as the base resolution of the input data.

This implied a compromise between the higher resolution of CLC and the lower resolution of interpolated meteorological data.

The FPI would then be derived with a 4.4 km spatial resolution. To match

the spatial resolutions, pixel values were replicated for the case of meteorological data, and a two-step generalisation was performed in the case of the fuel map, which was initially derived with the same resolution as the CLC, i.e. 100 m pixels.

5.3.2.1 Satellite data

At the Joint Research Centre (JRC) a satellite-tracking antenna receives daily NOAA AVHRR HRPT data. Within JRC, these images are pre-processed up to the computation of the NDVI.

Finally, the images are geometrically corrected, and an overall mosaic of Europe is created at a spatial resolution of 4.4 x 4.4 km.

Further information on process can be found in Kerdiles 1997). For the European FPI, historical NOAA –AVHRR mosaics processed at JRC were used in several ways: - First, daily cloud masks were extracted from the

NOAA images and used to derive a daily cloudiness percentage (%) grid for the study area.

- Second, 10-day Maximum Value Composites (MVC) of NDVI historical images were used to compute Relative Greenness (RG) (Burgan and Hartford 1993)

- Finally, the historical NDVI maximum value at a given location was used along with the overall historical maximum to parameterise the live-ratio values, as explained below.

Both the RG and the live-ratio were kept at the 4.4 km spatial resolution of the NDVI images from which they were derived.

5.3.2.2 Meteorological data

The moisture content of small dead fuels is considered the major factor in determining fire ignition.In this model, small dead fuels are represented by the Ten-Hour Time Lag Fuels.

The moisture content of small dead fuels (Fm10hr) is determined as a proportion of the equilibrium moisture content (Emc) of the surrounding atmosphere, as follows: (FOSBERG and DEEMING 1971)

Fm10hr = 1.28 x Emc

The Emc determines the amount of water vapour that a given piece of wood can hold. (SIMARD 1968).

In order to compute Fm10hr for a given day it is necessary to use values for the air temperature, relative humidity, precipitation, and percentage of cloudiness for that day.

The two last variables are needed to correct for solar heating and rainfall respectively.

Except for the percentage of cloudiness – derived from NDVI images - these meteorological variables were extracted from the MARS meteorological database located at the JRC (TERRES 1999).

Once this institution receives data at the station level, then theses data are spatially interpolated into 50x50 km grids according to the CGMS procedure.

This procedure takes the following characteristics into account: distance, difference in altitude, difference in distance to coast, and climatic barrier separation (VAN DER GOOT 1997).

The 50x50 km pixel size was imposed by the low density of meteorological stations, particularly in some areas of southern Europe.

More specifically, to compute the meteorological inputs of FPI three interpolated variables were extracted from MARS: maximum temperature of the day (°C), mean daily vapour pressure (hPa), and total daily rainfall (mm).

The daily rainfall is the sum of precipitation between 6 UTC on day D and 6 UTC on day D+1.

The interpolated actual vapour pressure and the interpolated maximum air temperature were used to compute the relative humidity - which was not directly available from the database- previous computation of the saturation vapour pressure.

EUFIRELAB

D-08-06.doc 25

5.3.2.3 Derived data sets

5.3.2.3.1 Fuel types map

In Europe, fuel type maps exist only at the national or local scales in some countries.

It was thus necessary to derive a European Fuel types map in order to be able to compute the FPI.

This fact implied the need for a method to link the ecological aspects of the forest fuels with the fire risk determinants.

The chosen approach required obtaining, from a land cover classification, a general framework for presenting the land cover regions.

Using GIS tools, these land cover regions were stratified into sub-regions according to phyto-sociological criteria, which accounted for the floristic composition and other factors governing the distribution of the vegetation.

Two databases were thus cross-referenced: the European classification Corine Land Cover (CLC) and the Map of Natural Vegetation of Europe (MNVE) (Conseil de l’Europe/Commission des Communatés Européennes 1987).

The CLC has a fairly good spatial resolution, with a minimum mapping unit of 25 ha.

However, thematically the CLC can be considered relatively poor, since only 44 classes are distinguished.

On the other hand, the MNVE has a lower spatial resolution (mapping scale of 1:3000.000) but depicts more than 100 Vegetation associations (V-A).

These associations delimit ecological areas that are relatively homogeneous, characterised by the predominance of natural or sub-natural primary vegetation.

First of all the CLC was used to mask out agricultural and not vegetated land.

Then the remaining land cover classes were subdivided into different spatial sub-regions according to the existing vegetation associations.

Next, the relationship between these Land-cover / V-A units and the standard National Fire Danger Rating System (NFDRS) (DEEMING et al. 1978)fuel types was investigated. Finally, the NFDRS fuel model key was then used to assign to each identified sub-region a fuel type depending on the characteristics of the predominating understory.

In this process there are several issues that are worth to stress: - The MNVE identification of the vegetation

associations is based on knowledge of the ecological factors that govern the vegetation distribution (mainly climate and soil). This information is also included in the MNVE data set along with the description of each V-A, and was also taken into account when assigning the fuel types with the help of the fuel model key.

- During this process the CLC is essential, because the MNVE depicts potential vegetation associations that sometimes do not coincide with existing ones. In this case the CLC is given priority and the fuel type is assigned directly to the land cover class.

Overlay of the two data sets was done at the CLC’s spatial resolution, resulting in a 100 m spatial resolution fuel map, which was subsequently generalised to 4.4 km resolution using a two-step procedure.

Notice that this approach has similarities with: - other global land cover/vegetation mapping

experiences in which ground truth data was not used, such as the 1-km Global Land Cover Characteristics database. (LOVELAND et al. 1991),

- the first phases of other experiences in which ground data was only used in the last phases of the labelling process. This is the case, for instance, of the NFDR Fuel

Model Map. (BURGAN et al. 1998)

5.3.2.3.2 Live-ratio map

Following the original FPI methodology, the maximum Live-ratio was initially obtained from a Fuel types map.

However, the building of such a map, particularly when fieldwork is not feasible, entails some subjectivity.

In order to address this problem, it was proposed (BURGAN, 1999) the use of a maximum live-ratio map derived through a re-scaling of NDVI historical maximum values in each pixel.

The maximum NDVI values were computed for a five-year interval 1994-1998), and the live-ratio was then derived as:

Live-ratio=0.25 + 0.50 x (NDVImax / NDVIabsolute-max)

Where - NDVImax represents the maximum NDVI for a given

location, - NDVI absolute-max is the overall maximum NDVI on

any location in the NDVI mosaic. The computed live ratio values are scaled between

0.25 and 0.75, in accordance with the live-ratio values defined for the NFDRS.

5.3.2.3.3 Extinction moisture map

In the FPI model the extinction moisture (ExtM) value of a pixel is deduced from the fuel type assigned to that pixel.

However, the ExtM content of broad vegetation types is normally found in the literature and so it was reasoned that the direct assignment of ExtM values was simpler, and could be closer to reality, than their indirect assignment through the Fuel types map.

Thus it was decided to derive an Extinction moisture map by assigning an ExtM value to each identified sub-region.

Table 6 shows the CLC classes considered in the analysis and the fuel types assigned to them.

The fuel type before the brackets indicates the general assignation.

Fuel types between brackets correspond to the different identified sub-regions (therefore they depend on the predominant vegetation association).

The fourth column shows the initially assigned ExtM values (those from the NFDRS).

Last column shows the modified extinction moisture values.

EUFIRELAB

D-08-06.doc 26

To obtain this table, several sources of information were consulted.

These comprised the extinction moisture values in the Behave and the NFDRS fuel models.

All the available documentation was consulted and the conclusions derived discussed with experts until the ExtM value was agreed.

These conclusions were used to revise the initially assigned ExtM values, which were only changed in case of contradiction with the findings of the experts.

Table 6 enables the comparison between the ExtM values initially used, and new ExtM values assigned when building the ExtM map.

It can be seen in this table that the use of the ExtM map simplifies the European FPI model.

5.3.3 Computation of FPI

Once the necessary inputs were derived, following the original methodology (Burgan et al. 1998), the Fire Potential Index was computed as:

FPI = 100 x (1-Fm10hcorrected) x (1-Lr)

Where Fm10hcorrected is the moisture content of the small dead fuels corrected by the solar heating and the extinction moisture content, and Lr is the Live-ratio corrected by the Relative Greenness, i.e. the part of the load that is live.

FPI values range 1 to 100, with 1 meaning “no fire risk” and 100 meaning “highest risk”.

For operational purposes this range has been simplified to 5 risk categories from very low to very high.

The FPI map calculated for the EU is avaliable at the web site

http://inforest.jrc.it/effis/viewer/viewer.html

5.4 LONG-TERM INDICES

Long-term fire risk indices are those computed from variables that do not change in a short lapse of time.

Consequently, these maps are updated with a yearly (or longer) frequency.

Two types of long-term indices could be included in the EFFRFS.

The first one provides the probability of fire occurrence, while the second one indicates the likelihood of damage to a forested area in terms of economic or environmental losses.

5.4.1 Probability of Fire Occurrence2.

This index intends to evaluate the probability of fire ignition, rather than the propagation of the fire.

Fire propagation is more related to the dynamic indices that have been presented in previous sections.

Aspect was included in the model as representative of the topographic conditions that may influence fire ignition.

It is related to the type and condition of the available fuels for ignition.

Since the influence of socio-economic factors (human factor) is difficult to model, fire recurrence was introduced in the model as a surrogate of these factors.

Fire statistics for southern Europe show that the probability of fire occurrence is higher in those areas that have historically suffer a high rate of fire incidence.

The method developed is based on multi-variant regression techniques.

Two sources of data of the European Commission have been used: on the one hand, the database of fires of the Main directorate of Agriculture; on the other hand, diverse data bases of the statistical office, EUROSTAT.

37 independent variables was extracted Of these two sources of data (see Table 7), including in three thematic fields: economic, social and environmental statistics.

All of the variables were chosen because of their likely relationship with the forest fires phenomena in the Mediterranean area as described in the literature on forest.

Some of them, like the fuel types and the topographic variables, account directly or indirectly for the hazard.

The others are deemed to somehow affect the probabilities of ignition.

The third thematic group would integrate the information relative to fuel types and the fisiographics variables; the rest would be socio-economic aspects.

In this model the dependent variable is Annual average of fires, weighted by province (or departement, in France).

The independent variables were not homogeneous, neither spatially nor temporally, for all the countries in the Mediterranean basin, and therefore it was decided to derive two models.

2Based on (Sebastián et al. 2001 and Sebastián, 2004)

EUFIRELAB

D-08-06.doc 27

The first model (Model 1) was built for the five EU Mediterranean countries with the set of 32 variables that ensure the homogeneity in the temporal and spatial resolution for the whole study area, i.e.: Portugal, Spain, France, Italy and Greece.

Other variables considered important in explaining the fire regime in the Mediterranean, such as surface of permanent grasslands, agrarian subsidies, and some economic statistics, were only available (at the required resolution) for Spain, France and Italy.

Hence, a second model (Model 2) was built for this reduced region using a total of 37 explicative variables.

Concerning the methodology, random sampling techniques were used to divide the original sample into two subsamples.

One was used to fit a linear regression model, and the other to validate it.

The original sample for Model 1 consisted of 323 observations (i.e. provinces), from which 194 were used to fit the model, and 129 to validate it.

The sample size for Model 2 included 245 observations.

From this sample, 156 observations were used to fit the model and 89 to validate the fit.

The sampling process was carried out 10 times for both models leading to 10 different and independent samples to fit the models and 10 corresponding subsets for their validation.

The variable selection was performed by searching for the best intermediate model (sets of explicative variables) for each of the obtained sub-samples.

The information derived from these regressions was then used to fit the final, definitive models.

The intermediate models were derived using robust regression, which provides suitable fits even when outliers are present.

An ordered list of the variables that resulted selected in the final Model 1 (all southern European countries) can be found in Table 8; and final Model 2 (Spain, France and Italy), in Table 9

An important consistency was found between the two models derived.

This supports the thesis of the selected variables as the better explicative variables of the forest fire phenomena at regional (Mediterranean) scale. The validation of the models was done through the analysis of their predictive ability.

Analysis of the spatial autocorrelation was performed to dextermine whether the selected models take into account the spatial structure of the dependent variable.

6 variables were coincident between the two models: two types of land cover (presence of shrub and mixed forest), two variables relative to the agrarian production (cattle and cereal production), a socio-economic variable (unemployment) and another fisiographical one (altitude).

Figure 20 presents an example of a fire probability map for southern Europe.

5.4.2 Likely Damage.

There are natural areas that are of particular interest.

This may be due to many different reasons, from the purely economic value of the timber to the unique environmental qualities of the area.

The aim of the likely damage index is to provide a method to highlight areas that should be strongly protected from forest fires.

In the Mediterranean region of Europe the most important quality of most forested areas is their environmental value.

Timber production is usually a secondary asset of these forests.

An added condition of these forests is the intrinsic difficulty for regeneration due to the lack of rain and the fragility of soils.

The likely damage was estimated by assigning to each cell a vulnerability degree.

The vulnerability index considers the likely damage that a fire can cause in a specific area.

This evaluation can be critical in areas of special ecological value, in susceptible zones of erosion or prone to the alteration of the hydric balance, and in areas close to human settlements.

This index considers three factors: - The potential erosion, obtained from the land cover

type, the slope and the regime of rains. - The protection level of a specific zone, that

considers its rareness nature, its fragility and its environmental interest.

- The distance to human settlements, that considers the human lives and the properties in danger.

To derive the mentioned variables, data from the Eurostat’s database were extracted and processed in a GIS environment.

All the long-term fire risks were normalized between the values 0 and 100.

This range was further divided into five fire risk classes from very low risk to very high risk.

An example of the likely damage index showing the levels of risk and the area for which long-term indices are computed is presented in Figure 21.

The two types of long-term indices (probability of fire and likely damage) could be integrated into a single index.

This so-called integrated long-term index would identify areas that are jointly subject to suffer forest damage and high potential losses.

The index would help forest fire services locate those areas to which the highest level of protection should be given.

EUFIRELAB

D-08-06.doc 28

5.5 FIGURES

Figure 19: Structure of FPI model calculation

Figure 20: Example of a fire probability map for southern Europe (San Miguel-Ayanz et al. 2002)

Figure 21: Example of a likely damage map for southern Europe (San Miguel-Ayanz et al 2002)

EUFIRELAB

D-08-06.doc 29

5.6 TABLES

Table 6: Proposed ExtM values for the considered CLC classes.

CLC Description Fuel type assigned

Live ratio Initially assigned Ext. Moist. (%)

New assigned Ext. Moisture

Permanent grassland L 0.67 15 15 Agriculture with natural vegetation T 0.67 15 20 Agroforestry areas L ( C ) 0.67 15 (20) 15 Broad-leaved forest R ( O,F, T ) 0.40 25 (30,15,15) 30 Coniferous forest T (D,H,C, P,A) 0.67 15 (30, 20, 20, 30, 25, 15) 25 Mixed forest R (O,P,C) 0.40 25 (30, 30, 20) 25 Natural grassland L ( A ) 0.67 15 (15) 15 Moors and heathland O 0.53 30 30 Sclerophillous vegetation F 0.60 15 15 Transitional woodland-scrub T 0.67 15 15 Sparsely vegetated areas A ( S ) 0.60 15 (25) 15 Inland marshes N 0.40 25 25 Peat bogs S 0.50 25 25 Salt marhes N 0.40 25 25

Table 7: Independents variables used in the study (Sebastián et al., 2004). Type Variable

Cereal production Wheat production Barley production Corn production Leguminous production Agrarian subsidies Agricultural total production Permanent grassland area

Agrarian production

Agricultural total area Goat production Bovine production Ovine production Cattle production

Addition goat and ovine productionPrimary sector Secondary sector

Economic Variables

Production by economic sectorsTertiary sector > 25 years old Unemployed index < 25 years old Population density Settlement > 20.000 inhabitants Protected areas

Social Variables Population variables

Road density Altitude Slope Fisiographic variables Orientation Land cover type 1 (%) Land cover type 2 (%) Land cover type 3 (%) Land cover type 4 (%) Land cover type 5 (%) Land cover type 6 (%) Land cover type 7 (%) Land cover type 8 (%) Land cover type 9 (%) Land cover type 10 (%) Land cover type 11 (%)

Environmental Areas

Fuel types

Land cover type 12 (%)

EUFIRELAB

D-08-06.doc 30

Table 8: Selected variables in Model 1 (Sebastian et al., 2001).

Variables in Final Model 1 Reg. Coeff.

Ln(Percentage of land occupied by Shrubs) 0.1195

Ln(Percent. of land occupied by Mixed Forests) 0.0973

Unemployment 0.0944

Ln(Surface of Permanent grasslands) -0.0936

Altitude -0.037

Ln(Percent. of land occupied by Broad leaves) 0.033

Total Cereal Production -0.0352

Total Cattle Production 0.0255

Table 9: Selected variables in Model 2 (Sebastian et al., 2001).

Variables in Final Model 2 Reg. Coeff.

Ln(Percentage of land occupied by Shrubs) 0.3775

Ln(Percent. of land occupied by Burnt areas) 0.2387

Total Cereal Production -0.2615

Ln(Percent. of land occupied by Mixed Forests) 0.2168

Altitude -0.2115

Total Goats production -0.1525

Ln(Perc. Land occup. by Mixed Forest- Agriculture) 0.1154

Total Cattle production -0.1095

Unemployment 0.1087

EUFIRELAB

D-08-06.doc 31

6 BASIC STRUCTURE AND CHARACTERISTICS OF THE PROPOSED EURO-MEDITERRANEAN WILDLAND FIRE RISK INDEX

6.1 GENERAL PRESENTATION

Fire risk assessment is routinely performed in most developed countries that are affected by forest fires.

Estimating forest fire risk involves identifying the potentially contributing variables and integrating them into a mathematical expression, i.e. an index.

This index, therefore, quantifies and indicates the level of risk.

A literature review of forest fire risk methods has shown how different approaches are used for the evaluation of fire risk.

The most relevant are producing updating evaluation of fire danger conditions over the Internet.

Most of these operational systems currently rely mainly on meteorological data (see section 2.3), which is considered the most critical variable for fire ignition and propagation, while other factors of risk are less considered or they are introduced in an experimental way.

Fuel loads and fuel water state is commonly considered, but the difficulties of obtaining accurate information on both variables are recognised.

Therefore, most commonly fuel loads are roughly estimated from land cover maps, and fuel state is associated to weather variation.

Additionally, human factors of fire risk are rarely considered, in spite of being considered a critical variable in fire ignition (MARTELL et al., 1987).

In addition to this lack of critical variables in fire danger estimation, another weakness of current fire risk assessment systems is related to the concept of risk itself.

Traditionally fire managers define fire risk considering the chance of starting or spreading a fire in a particular time and space.

According to FAO’s terminology (FAO 1986) forest fire risk is even more restricted, since is defined as “the chance of a fire starting as determined by the presence and activity of any causative agent”.

However, other natural hazards evaluate risk considering not only the potential occurrence of an event, but also –and more important in many cases- the potential damage which that event would cause on persons and values.

From this point of view, the traditional concept of fire risk, which is still active in most countries and regions, is not considering a critical component of risk, which affects the susceptibility to fire of the potentially affected region.

As it is well known nowadays, fire is a natural factor in many ecosystems, and therefore, from the ecosystem point of view, the impact of a given wildfire could be relatively contained, being within the limits of a normal natural disturbance.

Fire management should greatly benefit from having the possibility to consider in the decision making process the importance of the expected potential damage.

The assessment of the importance of the expected damage is closely related to the concept of vulnerability, which can be defined as the potential damage caused by the natural/technological disaster on persons and values (ecological, economic, etc.).

Consequently, an integrated assessment of fire risk should consider both likely occurrence of fires, as well as their potential damages (Figure 22).

The Eufirelab unit 8 intends to propose a coherent and integrated scheme for fire risk assessment, which would consider a wide range of risk variables, both associated to fire ignition and propagation, as well as to fire vulnerability.

Such a proposal will be based on previous experiences which have shown good potential for risk assessment.

As previously stated, most operational fire risk systems are focused on physical factors: weather data and fuel status mainly.

However, especially in Europe, a growing concern about the role of human activities in fire ignition and propagation is widely recognized (KALABOKIDIS et al., 2002; MARTÍNEZ et al., 2004; VÉLEZ et al., 2002).

Therefore, any risk assessment system should also include the consideration of socio-economic causes of fire, as well as the vulnerability associated to human beings (life, properties and values).

The proposed scheme will include these factors.

Another aspect to be taken into consideration to design an integrated risk assessment scheme regards the need of having consistent evaluation of risk at different temporal and spatial scales.

The great diversity of natural and socio-economic conditions strongly complicates this task, which is critical if risk conditions need to be objectively compared throughout Europe.

Obviously, any risk assessment system based on physical and human variables need to be adapted to local conditions, but a hierarchical design may help the use of a common scheme at different scales, and for a broad range of social and natural conditions.

On the other hand, Fire risk, as any other natural hazard, is not static, but needs to be updated regularly.

The temporal updating would strongly depend on how dynamic the different risk variables are.

Some would require a very short updating period (such as weather data), while others only need revising in few months-years (fuel types), or they can be considered static (elevation, slope).

Based on the indices revised in the preceding section and considering the availability of spatial information at the required scale (see previous deliverables of the unit), Eufirelab unit 8 will aim to define the scheme for fire risk integration.

It includes the definition of the risk variables, as well as the integration of those input variables in synthetic indices for fire risk mapping at European scale.

EUFIRELAB

D-08-06.doc 32

The system will include the assessment of Danger and Vulnerability (Figure 23).

The former is related to the probability of fire occurrence and propagation, while the latter is associated to the potential effects of fire on human beings and natural values.

The proposed system includes the estimation of two danger properties: ignition danger and propagation danger.

The former is related to the causal agents of fire (excluding lightning because no operational way to model this factor has been produced) as well as the conditions and properties of the fuel (live and dead), while the latter is associated to estimating the behaviour of the fire.

In this deliverable will be defined the two components of ignition danger (causal agents and fuel properties).

The propagation danger is related to fire behaviour (rate of spread and flame length).

The Danger component requires considering weather data, fuels and socio-economic data that should be derived from spatial analysis of point data (both measured and forecasted), as well as remote sensing images.

More details can be found in following sections.

As far as vulnerability assessment concerns, the main potential effects to be analysed have been divided into ecological and socio-economic factors.

Due to the small number of experiences using this component in forest fire risk index at global scale, a preliminary study of the variables to be included in a vulnerability index will be undertaken in this deliverable.

The socio-economic component of fire vulnerability will be focused on the potential damages on persons and properties.

The economic value (in different colour in the scheme) will not be addressed in this deliverable because the great difficulty to create this information at European scale.

The ecological component includes potential soil erosion after a forest fire and the environmental value of the territory.

The system will be designed in order to be implemented at European scale.

The target resolution will be of 1 km². The supra-national perspective, which will be

proposed for the EUMed countries, should potentially contribute to the improvement of the European Forest Fire Information System (EFFIS), currently run at JRC.

EUFIRELAB

D-08-06.doc 33

6.2 FIGURES

Figure 22: Integrated Fire Danger Rating (after ALLGÖWER et al., 2003)

EM - WFRI Propagation

Danger Index

Vulnerability Index

Fuel moisture

Population exposition

Environmental value

Economic value

Potencial Soil erosion

Probability of ignition (human factors)

Live fuels (satellite)

Dead fuels (meteo)

Human ignition

Ignition Danger Index

Figure 23: Scheme of the Euro-Mediterranean Wildland Fire Risk Index (EM-WFRI)

EUFIRELAB

D-08-06.doc 34

7 COMPONENTS OF THE FIRE RISK SYSTEM: IGNITION DANGER INDEX

7.1 FUEL MOISTURE

7.1.1 Introduction

The moisture content of fuel is a critical parameter in fire ignition because flammability is closely dependent on it (DIMITRAKOPOULOS & PAPAIOANNOU, 2001).

The burning capacity of the leaves and plants is inversely correlated to the moisture quantity.

For fire ignition danger, fuel could be divided in two categories: dead and live.

Dead fuels lying on the forest floor (fallen branches, litter, and foliage) are the most dangerous because they are drier than live fuels and more dependent on rapid atmospheric changes.

The moisture content of live fuels has a marginal role in fire ignition, but it is critical in fire propagation modelling because the amount of water is directly related to the rate of fire spread (CARLSON & BURGAN, 2003; SNEEUWJAGT & PEET, 1985; VIEGAS, 1998).

It is not easy to measure the water content of the leaves and plants in an operational way.

Two main directions are generally used: - water estimation from soil water quantity estimation,

using methods of water reservoir -the water quantity of soil is computed from rainfall values (input) and evapotranspiration (output)

- water estimation from remote sensing measurements of vegetation aspect. This first way is generally more or less integrated in the meteorological indices of fire risk.

Commonly, the estimation of dead fuel moisture content (FMC) is based on meteorological danger indices, which attempt to account for the absorption–evaporation relationships in inert materials (SIMARD, 1968).

The application of spatial interpolation techniques is required because meteorological data are frequently not available for fire prone areas.

Applying those indices to live fuel moisture trends could be more complex because live plants are much less dependent on atmospheric conditions than dead materials, given their mechanisms to extract water from the soil reserve and reduce evapotranspiration.

Recent research have obtained strong nonlinear relationships between live fuels moisture content and long-term meteorological codes (CASTRO et al., 2003; VIEGAS et al., 2001), but their results are species dependant. In this context, spatial interpolation techniques could introduce additional noise.

Within the context of fire danger estimation, remote sensing data have shown good correlations with live fuel moisture content (CHLADIL & NUNEZ, 1995; PALTRIDGE & BARBER, 1988).

Taking into account the previous comments, a combination between remote sensing data (related to live fuel moisture content) and meteorological index (related to dead fuel moisture content) will be proposed.

7.1.2 Live fuels (satellite information)

Remote sensing data have been frequently used to estimate the water status of plants, both in agricultural and ecological research (CARTER, 1991).

In the forest fire danger literature, the water content of plants is commonly expressed as fuel moisture content (FMC), defined as the percentage of water weight over sample dry weight:

FMC=[(Ww-Wd)/Wd)*100]

where Ww is the wet weight and Wd is the dry weight of the same sample.

This variable is mostly obtained through field sampling using gravimetric methods (VIEGAS et al., 1992).

Wet samples are: - weighed, then - oven-dried at 60 or 100 oC, and - weighed again to determine the dry weight.

FMC can be referred to for both live and dead species.

Within the context of fire danger estimation, good correlations between live FMC and multitemporal series of NOAA–AVHRR data have been found for herbaceous species using normalized difference vegetation index (NDVI) data (CHLADIL & NUNEZ, 1995; PALTRIDGE & BARBER, 1988), but problems were found for shrubs and trees (CHUVIECO et al., 1999a; LEBLON, 2001).

However, in the remote sensing literature, water content is usually expressed as the equivalent water thickness (EWT: water content/leaf surface), instead of FMC, because EWT is directly related to the absorption depth of the leaf.

Laboratory spectral measurements have been performed to estimate EWT, showing divergent results in the visible and near infrared (NIR) depending on whether they were done at leaf or canopy level, because of the indirect effects of water content changes on the whole plant (mainly through the modification of the leaf area index [LAI]).

However, short wave infrared bands (SWIR: 1.1–2.5 Am) have proven to be the most sensitive to EWT variations (BOWMAN, 1989; COHEN, 1991; DATT, 1999), although additional bands are required to reduce the uncertainty caused by other variables affecting SWIR reflectance.

Simulation studies based on radiate transfer models have recently identified a ratio of the near infrared (NIR) and SWIR band as the most appropriate for retrieving EWT at leaf and canopy levels (CECCATO et al., 2001, 2002), as previous experimental studies had suggested (HUNT & ROCK, 1989).

EUFIRELAB

D-08-06.doc 35

In spite of this progress, to estimate EWT from reflectance measurements, additional efforts need to be made to derive FMC from satellite data in fire danger studies because the amount of water per area is not as critical in fire propagation as the quantity of water per dry mass.

Assuming that the specific leaf weight (SLW= dry leaf weight/leaf area) is constant over time for single species, FMC may be considered a function of EWT (CHUVIECO et al., 2003a).

Still, when this relation changes significantly over time, FMC may be indirectly estimated as a result of the effects of plant drying on the decrease in leaf area index (LAI) values (mainly in shrub species) and chlorophyll content (herbaceous species).

Therefore, the estimation of FMC from reflectance measurements can be undertaken when the estimation is restricted to single or (physiologically) similar species.

This explains why strong empirical relations between FMC and satellite variables have been found by several authors (CECCATO et al., 2003; CHUVIECO et al., 2002; LEBLON, 2001).

Recent studies have shown that by better estimating other factors affecting canopy reflectance in the NIR and SWIR bands, particularly the leaf area index (LAI), it is possible to apply radiative transfer model (RTM) inversion techniques to obtain a reliable estimation of EWT and FMC (ZARCO-TEJADA et al., 2003).

Additionally, plant canopy temperature is affected by FMC changes because water availability is a critical parameter in plant evapotranspiration.

Based on this principle, several authors have tested the use of thermal images to estimate plant water content, mainly on crops (JACKSON et al., 1981; MORAN et al., 1994).

Forest and shrub canopies are more complex, but some workers have shown good relationships between the differences in air and surface temperature (ST) and fire danger hazard (VIDAL et al., 1994).

Because these differences are closely dependent on the density of vegetation cover, the combined use of surface temperature (ST) and NDVI have shown statistically stronger relationships with water content than either of the two variables alone (ALONSO et al., 1996; CHUVIECO et al., 1999a; PROSPER-LAGET et al., 1995).

7.1.3 Study case in Central Spain

The following study has been published in Remote Sensing of Environment (CHUVIECO et al., 2004b).

7.1.3.1 Objectives

The objective of the study case is the assessment of an empirical approach to estimate FMC of Mediterranean species based on multitemporal analysis of NOAA–AVHRR images.

The proposed method is built on statistical fitting of field collected FMC and satellite data, using a function of the day of the year to take into account the seasonal trends of FMC.

The empirical estimation was intended for operational retrieval of FMC in fire danger assessments.

Considering the current limitations of meteorological networks and fuel type maps, it was determined that the FMC estimation should not require external data sets other than the information derived from the AVHRR images and very simple vegetation maps.

The estimation was targeted at grassland and shrub species, which are the most dangerous in fire propagation of surface fires.

The empirical fitting was based on a long time series of field measurements of FMC for the Cabañeros National Park study site (Central Spain), but collecting field measurements at other sites with similar species validated it.

Previous work showed a strong statistical relationship between AVHRR-derived variables and FMC for the Cabañeros study site using only summer images (CHUVIECO et al., 2003a).

In this case, 2 years of field data were used for calibrating the model and 2 more years for validation in the same study site.

Additionally, strong relations were also found for Landsat-TM images (CHUVIECO et al., 2002) and SPOT-Vegetation images, showing consistent trends among the three sensors (CHUVIECO et al., in press).

This work follows the same trend towards finding consistent relations between FMC and satellite-derived variables, for operational use of satellite data in fire danger estimation.

In this case, the model is applied to spring and summer data, uses 4 years for calibration and 2 more for validation in the same study site, as well as five additional validation sites, located far away from the calibration area.

Additionally, it introduces a function of the day of the year to model the effect of the seasonal trends.

7.1.3.2 Methodology

One of the key elements to obtain a sound empirical estimation in remote sensing research is the availability of long time series of field data.

Field measurements must include valid samples of large areas, on which the satellite images will be acquired.

For this study, the Cabañeros National Park (located in Central Spain) was used as a calibration site because it offered unique opportunities for testing relations between FMC and satellite-derived variables.

First, the central area of the park is covered by grassland and shrublands on very gentle slopes.

Because it is a protected area, no agricultural practices are carried out, and therefore, temporal changes are associated to vegetation seasonal trends rather than crop alterations.

Fuel types sampled were grasslands (three plots) and several shrub species (two plots): Cistus ladanifer, Erica australis, Phillyrea angustifolia and Rosmarinus officinalis.

EUFIRELAB

D-08-06.doc 36

Plot sizes were 50 x 50 m, located 3 to 5 km apart, along a range of 20 x 5 km. A complete description of the fieldwork may be found in CHUVIECO et al. (2003a).

Field measurements were taken from early April to the end of September.

For the model calibration, a time series from 1996 to 1999 was used.

These samples were collected every 8 days. For validation purposes, another field campaign was

carried out in 2001 and 2002. In this case, the samples were collected every 16

days because previous analysis did not show shorter time changes in FMC values.

In each plot, three samples per species were collected. Average values per plot and period were computed.

Long-term trends of FMC values were compared for the different grassland and shrubland plots located in Cabañeros in order to test whether they were showing local or regional differences in FMC trends.

Considering the large distance between plots (3–5 km apart), if significant differences in average values between plots of grasslands or shrublands were not found, it could be concluded that FMC temporal changes of these two vegetation types are more significant than those changes caused by their spatial diversity.

Consequently, field measurements of small plots could be considered as representative of the temporal variation of FMC for large plots.

In this way, the field measurements could be soundly related to coarse resolution satellite images.

During 2001 and 2002, a different set of field measurements was taken in other regions of Central Spain for validation purposes (Figure 24): a site located in the province of Segovia was covered with grasslands; Atazar–Alberche with grassland and shrubs (a mixture of C. ladanifer and R. officinalis); Ibérica and Pre-pirineo with shrubs (R. officinalis and other shrub species), and Cádiz with C. ladanifer.

These plots are 200 to 500 km apart from the Cabañeros site and have different elevations but include similar species, parts of the Mediterranean ecosystem except of the Pre-pirineo site.

The plots were selected so as to include homogeneous plant coverage on as gentle slopes as possible.

However, the shrub species were frequently mixed, even with some trees.

Among the different mixtures, only those plots with a significant coverage of C. ladanifer or R. officinalis (more than 60%) were selected for validation purposes.

To assure consistency in the results, the field protocol of these assessment sites was the same as that of Cabañeros.

AVHRR images were acquired by the University of Alcala’s HRPT receiving station. Raw digital to reflectance conversion was based on NOAA coefficients (including degradation rates), and surface temperature (ST) was based on methods proposed by Coll & Caselles 1997).

Geometric corrections were based on orbital models and manual control points and automatic correlation improved multitemporal matching.

Daily data were synthesized into 8-day composites using maximum NDVI values.

The median value of a 3 x 3 pixel window was extracted from each composite and correlated against field measurements.

When comparing AVHRR images and field measurements, the potential noise caused by the great differences within the area covered may be reduced when using average values of species, instead of single plot averages.

For instance, average values of grasslands collected in a length of 10 km (three plots separated linearly 5 km each) would be a better representation of what an AVHRR pixel is actually measuring than single plot measurements.

Several authors have discussed the pros and cons of empirical and theoretical models in remote sensing research (STRAHLER et al., 1986).

Theoretical models have two main advantages: generalizing power and a better understanding of the parameters involved.

However, they are complex to generate because they require many input parameters that are often unavailable and are difficult to validate.

Empirical models are commonly based on statistical analysis.

They are simpler to formulate and provide a quantitative validation on their exactness, but they are difficult to generalize, especially when statistical relations are not based on physical properties.

In the field of water content estimation, a whole range of theoretical models has been proposed in recent years, most of them based on the radiative transfer function (BARET & FOURTY, 1997; CECCATO et al., 2001; CECCATO et al., 2002b; JACQUEMOUD et al., 1996; ZARCO-TEJADA et al., 2003).

They are solid approaches but require further assessment and must demonstrate their operational application with field campaigns.

These models estimate the EWT, which is the variable directly associated to leaf water absorption.

FMC is equal to EWT divided by SLW. EWT can be estimated using a radiative transfer function, but dry matter content cannot be directly retrieved because the water is masking its effect on reflectance (JACQUEMOUD et al., 2000).

For this reason, ZARCO-TEJADA et al. (2003) use a simplified inversion model to obtain dry matter for FMC estimation, after deriving EWT from a radiative transfer model.

Several authors (ALONSO et al., 1996; CHLADIL & NUNEZ, 1995; CHUVIECO et al., 1999a, 2002, 2003a; HARDY & BURGAN, 1999; PALTRIDGE & BARBER, 1988) proposed empirical fittings for estimating FMC from satellite data.

EUFIRELAB

D-08-06.doc 37

Most commonly, these studies were based on AVHRR images, although there are also some examples using Landsat-TM images (CHUVIECO et al., 2002).

The empirical model was based on linear regression analysis, where FMC was the dependent variable and the independent variables were AVHRR variables, NDVI and ST, and a function of the day of the year.

Two models were generated, one for grasslands and one for shrubs.

The major physiological differences between these two communities made it advisable to split the fittings.

Cistus ladanifer was selected as a representative of Mediterranean shrub species because it is widely represented in Spain.

Other species of the same family (Cistus sp.) are also broadly distributed across the Mediterranean basin.

The statistical model was built from 88 periods (22 periods of 8 days during 4 years: 1996 to 1999), covering spring and summer conditions of the Cabañeros National Park.

The time series include a wide range of rain patterns, with some dry years 1999 and 1997 with precipitation close to 200 mm in 6 months), and more humid ones 1996 and 1998), with 250 and 230 mm, respectively.

The equation was validated using data from the same study area (Cabañeros), as well as the other study sites previously described during the 2001 and 2002 spring and summer seasons.

Satellite variables considered in the linear regression were NDVI and ST.

The former would be positively related to FMC because the drying of the plant reduces chlorophyll activity in grasslands, as well as leaf area index in shrub species.

On the contrary, ST would be expected to be negatively related to FMC because the cooling effect of evapotranspiration is reduced when plants get dry and introduce mechanisms to reduce water loss.

Before obtaining the estimations of FMC from linear regression models, an analysis of trends between FMC and these two variables (NDVI and ST) was undertaken.

In a similar way to other study areas (ALONSO et al., 1996; KALLURI et al., 1998; MORAN et al., 1994; PROSPER-LAGET et al., 1995), NDVI showed a negative correlation with ST in both grasslands and shrublands for the spring and summer seasons.

This trend must be related to the physiological reaction of plants to higher temperatures and lower moisture contents, which, depending on the plants, may change leaf colour, deteriorate leaf structure, modify leaf angle distribution by leaf curling or reduce LAI by leaf loss, and/or decrease evapotranspiration.

Based on these relationships, some authors have proposed a regression model of NDVI and ST to estimate plant evapotranspiration (KALLURI et al., 1998) and fire hazard levels (PROSPER-LAGET et al., 1995).

A scatterplot of NDVI against ST for different FMC values observed in the Caban˜eros site showed the trend towards the appearance of low values of FMC when low values of NDVI and high ST values occur, both for grasslands and shrublands.

The trends are more evident for grasslands because they present a wider range of both FMC and NDVI values.

Following the logic of Verstraete and Pinty 1996), the design of an optimal index for discriminating different FMC values should be based on lines perpendicular to the main axis of NDVI and ST variation, which show potential sensitivity for discriminating FMC values.

Additionally, a temporal variable based on the day of the year (from 1 to 365) was included in the empirical fitting to take into account seasonal trends in FMC, following a logic already tested in Mediterranean conditions (CASTRO et al., 2003; CHUVIECO et al., in press).

Considering that these temporal trends are more contrasted for grasslands than shrublands, two different functions were computed:

FDg = (sin(1.5 x π x (Dy + Dy1/3)/365))4) x 1.3

FDc = (sin(1.5 x π x Dy/365))2+ 1) x 0.35

where FDg and FDc are the functions of the day of the year (Dy) for grasslands and C. ladanifer, respectively, and the sine angle is computed in radians.

This function was derived by fitting a periodical function to the temporal average of FMC values of grassland and C. ladanifer for 6 years of measurements in Cabañeros 1996–2001).

The function has a wider variation for grasslands than C. ladanifer, which agrees with the stronger contrast in the water content of herbaceous species.

The constant terms were used just to scale the functions in a similar range among them.

7.1.3.3 Results

For the 6 years of field data, the t tests applied to the temporal differences of the three grassland plots in the Cabañeros site, separated between 3 and 5 km, did not show significant differences among them.

Similarly, the average temporal trends of the two shrub plots located 3 km apart were not significantly different.

Therefore, it could be concluded that the average FMC values of small plots (50 x 50 m) are representative of large areas (several kilometres apart), at least in the Cabañeros site, and consequently, temporal data extracted from those small plots can be assumed representative of the large plots observable in AVHRR coarse-pixel size images.

As mentioned above, the equations to estimate FMC from AVHRR data were derived from multiple linear regression analysis, using NDVI, ST and FD for 4 years of Cabañeros field data 1996 to 1999).

EUFIRELAB

D-08-06.doc 38

The resultant equations were:

FMCg = -57.103 + 284.808 x NDVI – 0.089 x ST + 136.75 x FDg

FMCc = 70.195 + 53.520 x NDVI – 1.435 x ST + 122.087 x FDc

where: - FMCg and FMCc are the estimated FMC values of

grasslands and C. ladanifer, respectively; - NDVI is the normalized difference vegetation index

(range -1 to + 1); - ST is the surface temperature (in Celsius degrees); - FD, the function of the Day of the Year.

The determination coefficients (r2) obtained were 0.737 (p < 0.001) for grasslands and 0.672 (p < 0.001) for C. ladanifer.

Significance values of independent variables were lower than 0.01 for NDVI and FDg in the case of grasslands, and ST and FDc in the case of C. ladanifer.

ST was not significant for grasslands (p>0.5) because most of its discrimination power is included in the FDg variable.

However, it was incorporated to improve spatial estimations, given that FDg does not change spatially.

For the same reason, NDVI was kept in the case of C. ladanifer, in spite of its low significance (p = 0.187).

The contribution of NDVI to FMC estimation is more significant for grasslands because the drying process in herbaceous species is commonly followed by a loss of chlorophyll activity and a LAI decrease.

In both cases, the factor accounting for the seasonal trends (FD) is very significant (p < 0.001).

As expected, NDVI and FMC show a positive correlation, whereas for ST it is negative, confirming the physiological assumptions previously stated.

The assessment of these equations was carried out on other time periods (2001–2002) in the same study site (Cabañeros), as well as on other study sites, and good results were obtained in all cases.

Relations are very coherent at all sites, in spite of being at a distance of over 200 km for Segovia and Atazar–Alberche sites, and more than 500 km for the Ibérica site, and with different altitude ranges (up to 500 m of height increase in the case of Segovia).

For the Cádiz site, only four observations were available, but r2 values were also very high (0.96).

The Pre-pirineo site had no samples or neither grasslands nor C. ladanifer and was not included in this analysis.

The temporal trends are very well estimated, and deviations of actual versus predicted FMC values are low and have no consistent bias.

The worst estimation was observed for grasslands in late spring (early June) and in the middle of the summer (August).

The former is related to the sudden decrease in FMC, which in both years, changes from over 170% to just 30–35% in 16 days.

This severe decrease is reflected in the reduction of NDVI and increasing ST, but not as steep as the field FMC values show.

As a result, overestimations in this period reach up to 60% of FMC.

Additionally, negative FMC values were estimated in August of the second year, caused by very low NDVI values (below 0.1 for this period).

However, negative estimations are not a major obstacle for operational purposes because a simple filter could be applied to the empirical model to avoid them.

Additionally, FMC values of grasslands during most of July and August are below 30%, which may be considered the limit for live species.

Therefore, for practical purposes, grasslands may be considered as dead fuels for the central part of the summer.

The validation of the C. ladanifer showed an even better fitting than grasslands, with very close estimations both in spring and summer in the Cabañeros site.

The highest deviations from the observed FMC values never reached 20%, and for most periods, they are under 10% of FMC.

The other two validation sites for grasslands (Ávila–Segovia and Alberche–Atazar) also showed very good fittings, with r2 values of 0.881 and 0.905 (Figure 25).

There is a slight tendency towards overestimation in Alberche–Atazar and underestimation in Ávila–Segovia, but the relation in both cases is close to a 1:1.

The scattergram also shows a nonlinear estimation trend, especially in Alberche–Atazar, which may be related to the saturation of NDVI in the upper part of the range (BARET & GUYOT, 1991).

In fact, polynomial equations between observed and estimated FMC values provide r2 values higher than for lineal trends, with 0.96 for Ávila–Segovia and 0.95 for Alberche–Atazar.

Nonlinear relationships should also be explored at the calibration stage in the future.

FMC estimation for C. ladanifer shows good fittings in all assessment sites: Alberche–Atazar, Cabañeros and Cádiz, although for the latter, only 4 observations in the summer of 2001 were available (Figure 26).

The empirical model has a slight tendency towards overestimation, especially for lower values of FMC.

The lower values offer a better fitting between estimated and observed FMC with differences lower than 10% of FMC during mid-summer.

Considering certain physiological similarities between C. ladanifer and another widespread Mediterranean shrub, R. officinalis, the empirical function was also applied to other study sites where this shrub species had been sampled in the field.

The results were very positive for all sites (Cabañeros, Atazar–Alberche and Ibérica), with r2 over 0.85.

The temporal trends also show good fittings, with nonsignificant biases (Figure 27).

EUFIRELAB

D-08-06.doc 39

An underestimation was observed for the spring season, but the fittings improved in the summer, when fire danger is higher, and therefore, the need for accurate estimations is more demanding.

The absolute errors were higher for R. officinalis than for C. ladanifer, especially in the Ibérica site, which may also be caused by the mixture of the field plots, where several shrub species grow in the same area.

7.1.4 Probability of ignition related to live fuel moisture content

The empirical model generated in the previous section using NDVI, ST and function of Day of the Year showed a consistent predictive power to estimate FMC of grasslands and C. ladanifer, a typical Mediterranean shrub species.

The model was tested in plots located several hundreds of kilometers apart and with different altitude ranges.

Therefore, this model may be tested on operational scenarios in Mediterranean conditions, applying both spring and summer data.

The model only requires two basic satellite variables (NDVI and ST), the day of the year and a regional map of vegetation types, which distinguishes grasslands from shrublands.

This variable could also be derived from the multitemporal classification of AVHRR data, following any of the methods applied to derive global land cover maps (DEFRIES & TOWNSHEND, 1994).

Considering the spatial and temporal resolution of AVHRR images, the empirical index may be used for short-term estimations of fuel moisture content (Figure 28).

After estimating the FMC of live fuels through satellite information it is essential to transform these values to probability of ignition in order to be able to assess the probability of ignition in live fuels and to combine this probability with probability of ignition based on other factors (i.e. human factors etc).

To do so, CHUVIECO et al. (2004a) developed a simple method to convert FMC values to danger ratings based on computing ignition potential from thresholds of moisture of extinction adapted to each fuel.

Fire danger is restricted to the likelihood of fire occurrence, given a particular fuel moisture content.

This likelihood will be defined in terms of probability of ignition associated with fuel moisture content status.

These values can then be integrated with other variables associated with ignition sources (e.g., lightning, human), which could also be expressed in terms of ignition probability.

The method proposed from Chuvieco et al. (2004a), to convert FMC values to ignition potential associated with fuel status (IPf ranging from 0 to 1) is based on the concept of moisture of extinction (ME).

This is defined as the threshold moisture content above which a fire cannot be sustained (ROTHERMEL, 1972).

The ME of living fuels varies between 12% for some grass fuels and 200% for the needles of some conifer species.

For most live vegetative fuels the ME is in the range of 120%-160%, while for dead fuels it is in the range of 12%-40%.

This method assumes that ME values act as relative thresholds to ignition for each fuel, above which the IP dramatically decreases.

Although, the IPf for FMC values higher than ME would be zero, a conservative approach is recommended here, assuming that a marginal IPf exists even at high values of FMC.

For this reason, it is proposed to assign a maximum IPf value of 0.2 to the FMC that equals the ME value of each fuel.

FMC values lower than ME would have IPf values in the range of 0.2–1, the IPf being linearly inversely proportional to FMC values.

For FMC values greater than the ME, IPf values would range from 0.2 to 0.

Null IP (IPf = 0) was assigned to the maximum FMC value recorded in the historical series of FMC field measurements.

Schematically this method is based on the following algorithm (Figure 29):

If FMC > ME, then

IPf = {1– [(FMC – ME)/(FMCmax – ME)]} × 0.2

else

IPf = 0.2 + [(ME – FMC)/(ME – FMCmin)] × 0.8

where FMCmax and FMCmin are the maximum and minimum FMC values of each fuel type derived from field FMC samplings.

While these samplings are site specific, they might nevertheless be assumed to apply to fairly large regions with similar environmental conditions.

ME values are fuel-type specific. For dead fuels, ME values were taken from the

BEHAVE fire behaviour prediction system (Burgan and Rothermel 1984), ranging from 12% to 40%.

Grasslands with FMC values lower than 30% were treated as dead fuels and assigned ME values of 12% (model 1) and 15% (model 2) following values proposed for the BEHAVE model (BURGAN and ROTHERMEL 1984).

For live fuels, threshold values were taken from specialized literature.

For shrubs, an average value of ME was selected (105%), following the experimental results of DIMITRAKOPOULOS and PAPAIOANNOU (2001) and BURGAN (1979).

For annual grasslands, the ME threshold was fixed at 40% (ALBINI 1976).

According to ANDERSON (1982), the ME of grasslands is directly related to fuel depth.

This explains why short grasses have low values (12%–15%) and tall grasses (more than 70 cm high) have higher values (25%–30%).

EUFIRELAB

D-08-06.doc 40

The use of linear functions to convert FMC to IPf is based on the experience of several authors (such as DIMITRAKOPOULOS and PAPAIOANNOU, 2001) who have shown similar trends between FMC and ignition delay for a wide range of Mediterranean species.

The slope coefficient of the linear regressions was related in this case with the flammability of the species.

DELICHATSIOS et al. (1991) proposed a similar solution, they termed this regression coefficient “flammability delay increment.”

The use of a linear function instead of an exponential one (as suggested by other authors, such as TRABAUD 1976 and ALBINI and REINHARDT 1995) is justified by the fact that recent experimental results with numerous species of the Mediterranean Basin clearly show that there is a gradual, linear relation between flammability (as measured by ignition delay time) and moisture content (DIMITRAKOPOULOS and PAPAIOANNOU 2001).

Older studies used a limited number of species and not the ISO 5657-1986(E) methodology for measuring ignition time of materials (TRABAUD 1976; VALETTE 1992).

In a highly exponential (nonlinear) relationship, small variations of the FMC in the upper scale of the curve (i.e., near the threshold of the moisture of extinction) may result in relatively big differences in the flammability of the fuels.

However, Mediterranean species seem to respond with gradual fluctuations in their ignition delay time to changes in the FMC status (DIMITRAKOPOULOS and PAPAIOANNOU 2001).

This is especially noticeable in live fuels and highly flammable species (i.e., species with relatively low ignition delay time).

Efforts are underway to improve the highly exponential relationship between the moisture and fuel flammability in the Canadian Forest Fire Danger Rating System (MCALPINE, 1995; in BEMMERZOUK, A.M., 1997)

7.1.5 Probability of Ignition in Dead Fuels (Meteorological Index)

7.1.5.1 Meteorological Danger Indices

Dead fuel includes a wide range of materials (senescent grasses, dry leaves, small twigs, and organic material in the topsoil).

Water content of dead fuels is the most important factor in determining fire danger potential.

On one hand, the water content of the fuel is inversely related to the probability of ignition, due to the fact that part of the energy necessary to start a fire is used up in the process of evaporation right before the fire starts to burn (CHANDLER et al. 1983).

On the other hand, water content also affects fire propagation since the source of the flames is reduced with humid materials, therefore reducing flammability (ROTHERMEL 1972).

For this reason, the estimation of water content in these fuels is a critical variable both in determining fire ignition in a specific time and place, as well as in predicting the behaviour of the fire itself.

The water content of dead fuel is constantly changing, depending mainly on atmospheric conditions (SIMARD 1968).

Loss or gain of water content will vary depending on the physical and chemical characteristics of the fuel and the presence of varying atmospheric activity (rain, condensation, etc.).

The water content of dead fuels is determined by various methods.

The most precise one is direct sampling by gravimetric methods.

Following this approach, water content is computed from the difference of wet and dry weights of the samples.

Most commonly, the water content is expressed as a percentage of the dry weight (BLACKMAR and FLANNER 1968; DESBOIS et al. 1997).

Direct sampling provides exact measurements, but it is costly and labour intensive, especially when wide area estimations are required.

Additionally, this method does not provide an instantaneous measurement, as the samples must be oven dried during a certain number of hours (24 or 48 have been commonly suggested).

Other methods are based on the use of previously calibrated wooden sticks that are assumed to be good representatives of certain types of fuel (SIMARD 1968).

In the USA, standard 10-hour fuel sticks with an oven-dry weight of 100 grams are commonly used.

The sticks are continuously weighed.

Finally, meteorological danger indices (MDI) have a long tradition in fire danger estimation, because they comprise different critical variables related to fire ignition and fire propagation.

The meteorological danger indices vary in complexity and in the number of variables to be considered, from those that only require temperature and relative humidity to those which are based on complicated numerical models (for a summary of the models most used see VINEY 1991).

These indices rely on current and past weather conditions, since they also try to estimate the degree of dryness of different forest fuels.

Since most of the countries have a relative dense network of weather stations for different purposes (agriculture yield prediction, disaster prevention, traffic regulation, etc.), MDI values can be operationally computed for extended territories.

Additionally, with the growing availability of automatic weather stations, these indices may be computed very frequently and measured in real time.

However, often the location of weather stations is not very appropriate for fire danger estimation, since they are located in urban or agricultural areas.

Therefore, spatial interpolation techniques to estimate weather variables at forested areas are required.

These interpolation methods always introduce a certain estimation error, which is added to the actual estimation of fuel moisture content (FMC).

EUFIRELAB

D-08-06.doc 41

Even so, this method is the most widely used to determine the FMC of dead fuel.

There are a lot of fire danger meteorological indices which use in different ways air or/and ground temperature, air or/and wood sticks humidity (relative or absolute), wind speed, and soil water reserve as an expression of the possibility for the plant to get a water satisfaction rate.

From all these meteorological indexes found in fire prevention literature, the Fine Fuel Moisture Content (FFMC) of the Canadian CFFWI (Canadian Forest Fire Weather Index) and the 10-h moisture code of the National Forest Fire Danger Rating System (NFDRS) provide a sound estimation approach for short-term changes in dead fuels FMC.

Both the CFFWI and the NFDRS have several components and their usefulness had been validated for fire prevention in several Mediterranean contexts (RODRÍGUEZ Y SILVA 2002; SEBASTIÁN 2002).

The FFMC estimates the water content of the top layer on the ground (L-layer) from measurements of temperature, relative humidity, wind velocity, and precipitation registered in the last 24 hours (VAN WAGNER 1987).

This water content index has an empirical character derived from the relationship between these meteorological variables and the water content of a standard fuel (jack pine and lodgepole pine type), at the same time integrating the accumulative effect of atmospheric conditions in the hours previous to the measurements.

This index has a timelag of approximately 0.66 days.

On the other hand, the index 10-h, which is part of the NFDRS, estimates the water content of combustibles with a width of 1.2 to 2.5 cm (BRADSHAW et al. 1983).

This index uses the concept of Equilibrium Moisture Content (EMC).

The EMC of a fuel element under given environmental conditions is the moisture content that the element will attain if left for sufficient time in those (constant) conditions.

The EMC is a function of the temperature and the relative humidity (SIMARD 1968), as well as the atmospheric conditions present at the moment of measuring the samples.

A study undertaken in the Department of Geography at University of Alcalá has compared the usefulness of two meteorological danger indexes in fire prevention used in Canada (FFMC) and the United States (10-h) in order to evaluate the FMC of dead fuels in a Mediterranean territory.

Taking measurements in the field over a period of 6 years compared the codes.

The results did not show a significant difference between the two; therefore it is recommended that the index with less meteorological variables (10-h) be used.

This index allows an estimation of FMC of flammable dead fuels with an RMS around 5% (AGUADO et al, in preparation).

As noted before, one of the main challenges of using these meteorological fire danger indexes is extending the measurement obtained in the specific weather stations to cover the rest of the territory (FUJIOKA 1987).

This problem is compounded more and more because the Forest Services use the GIS technology in forest fire prevention when they integrate spatial variations in the data.

As a result, once an index with a better predictive capacity has been selected, it is necessary to get a spatial distribution of this index.

The spatial dimension of meteorological indices can be obtained using a mapping methodology based on logical relationship between the value of an index (or of its components) and the environmental parameter values, as it has been explained in a previous deliverable (D-08-05).

This methodology improves the spatial information: from a map built with some points (the index value of which is indicated) and many empty cells, we get a map with all the cells filled with the index value for each cell.

Next section shows an example of this methodology at local scale.

7.1.5.2 Spatial interpolation techniques: the case of Alpes-Maritimes (France)

Mapping the meteorological indices values is in fact one of the most important steps of the risk determination.

As explained in previous deliverable D-08-03, meteorological indices are always computed from stations network data.

Every data is representative only of the sensor that measured it, and then there are two concepts concerning the use of meteorological data measured in a given place, to elaborate a meteorological risk index: - Either a region is covered for instance by a 20

stations network, and the index is computed for each station. That means the risk is known only for 20 points, and is unknown for all the places, which exist outside these points. The users, as foresters or firemen have to fill themselves, with their imagination, the numerous places where the value is not computed, and in fact they interpolate in their mind the absent values. This risk concept is a punctual one and the problem is the needed information is absent on large surfaces.

EUFIRELAB

D-08-06.doc 42

- Or, in the same region covered by 20 stations, it is considered that there are sub-regions constituted by different areas, which are relatively homogeneous. For example, in France, MeteoFrance decided they are 7 different « fire regions » in the department of Alpes-Maritimes (Nice). For every « fire region » (their surfaces are different from the others), the meteorological risk index is computed only from the stations, which are inside the fire region, generally 2 or 3 stations. In fact, each fire region is «virtual» because each meteorological variable which is used and contribute to compute the index of each fire region is an average value computed from the stations. Thus the index is obtained from data characterizing a non-existent place which results from computed values (means, barycenter from 2 or 3 stations).

In this case we can say that each fire region (from 200 to 1500 km2) index is supposed to represent an area which is homogeneous enough, so that in details, the risk is he same in each point of the region.

Obviously, especially in mountainous relief, as very often in Mediterranean regions, this homogeneity does not exist…

We propose here another possibility, that can be used everywhere, if a stations network exists: it consists to obtain with scientific arguments the most probable meteorological risk index value for a very small area, a pixel, the surface of which is for instance from 2500 m2 (50 m side) to 1 km2 (instead of some hundred or thousands kilometres…).

The principle is to interpolate data from a station to elaborate the risk value of each pixel.

In the previous deliverables D-08-03 and D-08-05, we explained the « environmental » regression methodology: a statistical law is found, establishing a link between the meteorological data of a place (a pixel) and the environment of this place, especially the topographic and relief environment (altitude, relative altitude, aspect, slope, sea distance).

Realizing index risk maps induce to answer to 2 sorts of problems and to make choices: - To decide the width of each pixel, in a raster GIS; - To decide which methodology will be used:

interpolation of data from the network stations and computing the index for each pixel, or computing the index for each station (from its data) and interpolation of the index for each pixel.

The meteorological index we have used in this work is I85/90 CARREGA index for 3 reasons: - We know it very well as one of us is the author! - This index is simple because composed with only

3 parameters (wind, air humidity, soil water reserve), that allows to understand more easily how maps are changing from a situation to another, and how spatial interpolation works;

- This index has been used by French firemen and foresters of the Alpes-Maritimes during more than 10 years 1990-2000) inside of an Expert System elaborated by Ecole Nationale Supérieure des Mines de Paris, and has proved to be very efficient, according to the users.

7.1.5.2.1 The width of each pixel: spatial resolution.

The smaller a pixel is, the more accurate the risk definition will be, and that may be very interesting, especially in complex topography.

But such a precision needs computer resources.

Working on a large area as Mediterranean Europe for instance, at a given moment, does not necessitate having a very high spatial resolution.

But, a very high resolution is needed for operational work, i.e. by foresters or firemen.

In this case, the spatial scale used has to be very accurate.

In this work we used a DTM of Alpes-Maritimes (4400 km2) with 50 m spatial resolution, (degraded to 1 km2 if needed).

The map of figure 30 gives an idea of the topography of the region where this methodology was checked (department of Alpes-Maritimes).

In fact, we decided to work on the south part of this region because excepted one, all the available automatic stations are located in this south part, allowing thus to obtain a higher measurement network density for meteorological data.

As explained in previous deliverables, different environmental parameters can be obtained from the DTM, as: - Altitude of each pixel; - Relative altitude, indicating if a given pixel is at a

lower position (bottom of a valley) or dominating; - Aspect, with a value varying from 1° (North) to 180°

(South) that is the maximum, East and West being the same;

- Slope, above the given pixel, in degrees; - Sea distance.

One example of maps is given that allow comparing two parameters as aspect, for 50m pixel (fig. 31) and 1km pixel (fig. 32).

7.1.5.2.2 Two different logics: direct interpolation of the index, or interpolation of meteorological data in order to compute the index.

Direct interpolation of the index, 50 m definition

It is in fact the simpler way to obtain a meteorological risk index map.

The methodology needs a station network and an environmental regression model.

We decided to use always the same model, even if according to the date and the corresponding meteorological situation, the score of the regression analysis is not the same.

The environmental parameters used are: altitude (m), relative altitude (m), aspect (°), slope (°), distance from the sea (km).

EUFIRELAB

D-08-06.doc 43

Maps of figure 33 and figure 34, at 06h and 12h UTC are risk index interpolation maps using a DTM 50 m resolution for each pixel.

The risk index value of each 50 m pixel was interpolated directly by environmental regressions, from the risk index value computed for each meteorological station.

On 28 August at 6:00, the influence of altitude is obvious at the end of the night: risk is higher on slopes facing to South but especially on higher altitudes.

High risk values (red) are not numerous, and the dominant situation is a low risk value, because we are just at the beginning of the day, with generally low wind speed, and high air humidity.

The high value of the index in altitude is due to wind speed.

At 12:00, the difference is obvious, due to decreasing air humidity (temperature has increased) and increasing wind speed (due to air heating and turbulence).

Interpolation of the meteorological variables which constitute the risk index value, 50 m definition

In this case the interpolated values are not the risk index initially computed for each station.

The interpolation concerns the meteorological variables.

These ones are interpolated from the station network for each pixel of the map, taking into account its environmental parameter.

Obtaining its meteorological values allows computing for each pixel its risk index value.

The environmental parameters used to interpolate meteorological data are also: altitude (m), relative altitude (m), aspect (°), slope (°), distance from the sea (km).

Maps of figure 35 and figure 36, at 06h and 12h UTC are risk index maps resulting from meteorological variables interpolation (using a DTM 50 m resolution for each pixel).

The risk index value of each 50 m pixel was computed directly from its meteorological values.

These values were interpolated by environmental regressions, from meteorological values of each meteorological station.

Comparing the two maps (interpolation of the risk index or interpolation of meteorological components of the index, and then computing the index) allows noticing differences of behavior: mapping directly the index tends to smooth the local differences and prevents to know what is the meteorological variable which dominates the others.

On the opposite, the risk map resulting from interpolation of meteorological variables shows easily some “spots” the risk of which is obviously different (blue spots of fig 35 and 36).

They are due to the interpolation method of wind speed which is not an environmental one because the logic of spatial variation of wind does not obey principally to topographic conditions.

Thus the interpolation can be made by complex models or by spatial interpolation as kriging or simpler, as IDW method (weight of each point is related to the distance between other points).

Direct interpolation of the index, 1 km definition

The methodology is the same as section i), but the geographical scale changes: each pixel is 1 km instead of 50m.

Comparing both maps of fig. 33-34 and fig. 37-38 shows the differences concerning the spatial accuracy, which is less operational and usable when pixel width is 1km.

But the advantage of this 1 km dimension is that less computer capacity and memory storage is used.

Maps of figure 37 and figure 38, at 06h and 12h UTC are risk index interpolation maps using a DTM 1 km resolution for each pixel.

The risk index value of each 1 km pixel was interpolated directly by environmental regressions, from the risk index value computed for each meteorological station.

Figures 39 to 43 are 1 km definition risk index maps (index interpolated) for August 31st 2003.

The sequence of maps allows to evaluate the the risk level which is changing very quickly in time, and which is different at the same moment from a place to another one.

During this day a Foehn wind effect occurred at the end of the morning, with high temperature, very low air humidity and increased (but moderate) wind speed, with a soil water reserve, which was very low.

A “mixed fire” (forest, crops and houses) started about 12:00 UTC very close to the sea, burnt more than 200 ha despite of about 800 firemen and destructed 10 houses.

This fire was studied and explained in D-08-02. It is interesting to notice that the meteorological risk

index was very high in the place where the fire began (in dark at 12:00).

7.1.5.2.3 Conclusion

The possibility to obtain a map with a spatial repartition of risk indices, and especially of a meteorological index, seems to be very interesting, compared with the simple knowledge of some points only, or to large areas in which the risk value is supposed to be the same.

The methodology used influences the results as the comparison between the two sorts of maps shows.

But the most satisfactory way to obtain a map (interpolation of meteorological variables) from an intellectual point of view is also the longer.

Thus, it is easier to interpolate directly the value of the risk index.

In fact the validity of the result depends upon the performance of the coefficient of correlation.

This one is generally better for water soil reserve than for humidity, and better for 50 m definition than for 1 km.

EUFIRELAB

D-08-06.doc 44

7.1.6 Figures

Madrid

Madrid

Sevilla

Barcelona

Valencia

BilbaoCoruña

1

2

Zaragoza

1 Cabañeros (700m)2 Atazar (900m)3 Alberche (800m)4 Segovia (1200m)5 Ibérica (900m)6 Pirineo (1000m)7 Cádiz (400m)

43

5

6

7

Grasslands

Shrublands

Both

Name (Elevation)

Madrid

Madrid

Sevilla

Barcelona

Valencia

BilbaoCoruña

1

2

Zaragoza

1 Cabañeros (700m)2 Atazar (900m)3 Alberche (800m)4 Segovia (1200m)5 Ibérica (900m)6 Pirineo (1000m)7 Cádiz (400m)

43

5

6

7

Grasslands

Shrublands

Both

Name (Elevation)

Figure 24: Location of the sampling plots

Grasslands Ávila-Segoviay = 0.9525x - 3.8232

R2 = 0.9053

0

50

100

150

200

250

300

350

0 50 100 150 200 250 300 350

Obse rv e d

Est

imat

ed

Grasslands Alberche-Atazar y = 0.841x + 6.908

R2 = 0.8815

0

50

100

150

200

250

300

350

0 50 100 150 200 250 300 350

Observed

Est

imat

ed

Figure 25: Observed and estimated FMC values for grasslands in Alberche-Atazar and Ávila-Segovia

Validation sites (2001 and 200)

EUFIRELAB

D-08-06.doc 45

Cistus all sites y = 0.7376x + 29.912R2 = 0.7241

020

4060

80100

120140

160180

0 20 40 60 80 100 120 140 160 180Observed

Estim

ated

Cabañeros Cádiz Alberche-Atazar

Figure 26: Observed and estimated FMC values for C. ladanifer in Alberche-Atazar, Cabañeros and Cádiz. Validation sites (2001 and 2002)

Rosemary Cabañeros (2001-2002)

020406080

100120140160180200

3-4 29-5 16-7 2-9 9-5 27-6 13-8 30-9

FMC

Observed Estimated

Figure 27: Observed and estimated FMC values for R. officinalis in Cabañeros.Validation sites (2001 and 2002)

Figure 28: Live fuel moisture content map from Euro-Mediterranean countries (August, 10th; 2004).

Based on CHUVIECO et al., 2004a, processed by A. CAMIA, JRC

EUFIRELAB

D-08-06.doc 46

FMC PIf

00.20.40.60.81

max

150

100

50

0

ME

FMC PIf

00.20.40.60.81

max

150

100

50

0

ME

Figure 29. Scheme to convert fuel moisture content (FMC) to ignition potential (IP) (example of dead fuels)

Figure 30: The topography of the Department of Alpes-Maritimes: altitude.DTM with Pixel width: 50 m.

EUFIRELAB

D-08-06.doc 47

Figures 31 and 32: Aspect of each 50 m pixel (left) and 1 km (right)

Figures 33 and 34: High definition (50m) Meteorological risk index map on 28 August 2003 at 06 and 12 UTC.

Figures 35 and 36: High definition (50m) Meteorological risk index map on 28 August 2003 at 06 and 12 UTC.

EUFIRELAB

D-08-06.doc 48

Figures 37 and 38: Low definition (1 km) Meteorological risk index map on 28 August 2003 at 06 and 12 UTC.

Figures 39 and 40: 1km definition risk index maps for 31st August 2003 at 03:00 and 06:00 UTC

Figures 41, 42 and 43: 1km definition risk index maps for 31st August 2003 at 09:00, 12:00 and 15:00UTC

EUFIRELAB

D-08-06.doc 49

7.2 PROBABILITY OF IGNITION (HUMAN FACTORS)

7.2.1 Introduction

Human activity is the main factor of forest fires ignition worldwide, having a special impact on Mediterranean countries of Europe.

From different statistical sources we know that most of the wildfires in Mediterranean Europe (above 90%) occur as a consequence of human activities that can directly act as fire ignition sources or indirectly create the conditions that favours fire ignition and/or fire propagation.

For example, according to data published by the Spanish Forest Service (DGCN) for the period 1988-1999, 96% of the fires in Spain were produced, either directly or indirectly, by human causes, which indicate the close link between fires and human activities.

The only natural cause of fires is lightning, which in Spain accounts for a small percentage of only 4% according to statistics of the past 15 years.

This is probably due to the fact that storms are not as important in Spanish forested areas as in other parts of the planet.

In fact, this may well be the only meteorological phenomena in this climatic area that does not favour fires ignition.

Curiously enough, this distribution of fires according to its cause is quite different to other typical Mediterranean ecosystems such as the Californian one, where lightning may cause up to 40% of fires.

There are other non Mediterranean parts of the planet where lightning is of high importance such as in Scandinavia and Russia (20 to 30%), Western North America (40 to 60%) and in the Australian region of Victoria (10 to 30%).

Consequently, it is clear the importance of including human factors in any comprehensive fire danger index.

Nevertheless most current operational indices of fire danger rely on physical variables (mainly weather data), and human factors are not ordinarily considered (Chuvieco et al., 2003b).

Modelling human factors associated to fire ignition is very difficult, since economic and recreational activities linked to fire are very disperse in time and space.

An alternative to model human factors of fire danger is the estimation based on indirect indicators that relate fire occurrence with spatial variables linked to human activities.

Human ignition danger may be defined as the probability of a fire occurring as a result of the presence and activity, either directly or indirectly, of human beings.

The evaluation of human activity as an agent of fire ignition is a complex task, since data on human activities on forested areas are rarely available (VEGA GARCÍA et al. 1995).

In fact, the temporal and spatial data required to evaluate the human factors in fire danger rating simply do not exist (MARTELL, OTUKOL, and STOCKS 1987), while other variables of fire danger such as temperature or relative humidity are routinely generated.

Moreover, human activities are very dynamic on time and space, which difficult the estimation of specific spatial patterns, or rather these are more difficult to determine, as in the case of pyromania or specific deliberate motivations and attitudes.

However, for other causes, such as those relating to recreational activities or agricultural burnings are easier to model using spatial variables.

The complexity of dealing with the human factors of fire danger have frequently led to investigators and fire managers to either leave them out of their risk models, or consider them only marginally.

Most commonly, when those factors are considered, the estimations are based on indirect assessments of human risk activity, namely indicators of activities which are the usual cause of fire and which are generally structural in nature, i.e. population density or urban-wildland interface.

The first studies on human factors of fire danger were based on indirect variables, obtained mainly from censuses and survey sources (ALTOBELLIS 1983; BAIRD 1965; BERTRAND and BAIRD 1975; CHRISTIANSEN and FOLKMAN 1971; COLE and KAUFMAN 1963; DOOLITTLE 1979; DOOLITTLE and WELCH 1974; FOLKMAN 1965, 1973; JONES, TAYLOR, and BERTRAND 1965).

Later on, in the eighties the human factor is analysed spatially, and it is frequently based on cartographic aspects, that considered some human variables along with natural features, such as slope or fuel types (Aerial Information Systems Inc. 1981; CHUVIECO and CONGALTON 1989; DONOGHUE and MAIN 1985; DONOGHUE, SIMARD, and MAIN 1987; LYNHAM and MARTELL 1985; PHILLIPS and NICKEY 1978; MARTELL, BEVILACQUA, and STOCKS 1989; YOOL et al. 1985).

The number of such studies has increased in the last years due to the greater availability of digital cartographic and statistical information managed through Geographical Information Systems (GIS) (ABHINEET et al. 1996; CHOU 1992; CHOU, MINNICH, and CHASE 1993; CARDILLE, VENTURA, and TURNER 2001; CHUVIECO and Salas 1996; CHUVIECO et al. 1999b; LEONE et al. 2003; SALAS and CHUVIECO 1994; THOMPSON 2000; VASCONCELOS et al. 2001; VEGA-GARCÍA et al. 1995; VEGA GARCÍA et al. 1996).

These recent studies commonly use variables related to recreational activities in forested areas, proximity to roads and trails, population density, distance to human settlements, forest property types, etc.

These variables are easier to spatialize, but they may have marginal importance in some areas (VEGA-GARCIA et al., 1995).

Despite the importance of these empirical works, there are still many aspects of the fire-human relationships that require further research.

EUFIRELAB

D-08-06.doc 50

For instance, we do not have yet a clear understanding on the spatial consistency of those relationships between fire and human factors.

Some variables may be critical in some regions, while marginal in others, depending on the socio-economic and environmental structure of the territory.

Additionally, the human factors require a comprehensive assessment, which includes not only spatial-variables (distances, fragmentation, interfaces) but also socio-economic ones (unemployment, rural population age, population density…).

Recent conferences in forest fire research have emphasised the importance of this holistic approach (Viegas 1998, 2002).

In the following sections two examples of human ignition danger assessment are showed; one case at regional scale and another one at European Scale.

We are aware that both aproaches need to be improved and adapted in order to integrate them properly in a euro-mediterranean risk index using the previously specified resolution of 1 km2.

Finally a fire occurrence map is presented as a possible alternative

7.2.2 Estimation of human ignition danger at regional scale: the case of Alpes-Maritimes (France)

Natural ignition due to lightning in the Alpes-Maritimes (SE France) is relatively rare since most storms are usually accompanied by rain due to the proximity of the sea which induces conditions of high atmospheric vapour.

Approximately 90% of fire ignitions between 1991 and 2003 in the Alpes-Maritimes were of human origin, and this figure is probably an underestimate of long term averages since the data include the fires of 2003 which was a particularly hot and dry summer with more “dry” storms than usual during the month of August.

It is therefore reasonable to attempt to estimate the probability of fire ignition as a function of the location with respect to human settlements since it is in proximity to human activities that most fires find their origin.

7.2.2.1 Description of study area and database

Site location: the department of the Alpes-Maritimes was chosen since it includes both low altitude climatic conditions and Mediterranean vegetation with a dominance of summer fires and a high altitude mountain area where fires tend to occur at the end of winter.

The surface area of the department is about 4300 km2.

Data period: a 13 year period, from 1991 to 2003 represents the time during which data on fire ignition were systematically collected and recorded.

Description of the data: the French National Forest Office (“Office National des Forets”, or ONF) provided the bulk of the fire data for the period defined above.

This included information on the location and likely cause of fire ignition, as well as surface area burned (not used in this study).

The “Prométhée” database provides less comprehensive information for fires going back to 1973.

Fire ignition data were complemented by a Digital Elevation Model (50 m resolution), digitized road network from the “BD Carto” data base of the IGN (National Geographic Institute), and vegetation and land-use maps at a 1:25,000 scale from the Corinne Land Cover data base.

Elevation, vegetation, land-use, and road networks maps for the Alpes-Maritimes are shown in Figures 44 to 47.

In summary, the Alpes-Maritimes department is particularly mountainous in the North of the department and hilly in the South along the coast.

The Mediterranean climate has a pronounced dry period in the summer, which is accompanied by dry periods in the mountains during the winter.

Within a 10-15 km swath along the coast, the vegetation is typically Mediterranean, especially West of Nice.

The population of about 1 million inhabitants is concentrated in the coastal area with a succession of large cities (Cannes, Antibes, Nice, Monte-Carlo...).

The area of Mediterranean vegetation is therefore concentrated in the suburban fringe just North of the urban coastal area where houses are mainly individual villas surrounded by forests

Forest fire ignition risk can be considered high in this sector due to abundant dry vegetation in the summer and a relatively dense road network and a concentration of human activities.

Forest fire ignition: During the 1991-2003 period, a total of 550 fires were recorded in the Alpes-Maritimes.

A precise ignition location is available for only 362 (about 66%) of these, and a known can be defined for only 225 fires, or 50% of the total number of fires (Figure 48).

The spatial distribution of fire ignitions in the department according to known cause is shown in Figure 49, identifying 8 categories of fire ignition.

The remaining fires were classed as cause unknown

7.2.2.2 Modelling fire ignition risk : a first « intuitive » approach.

An initial approach to estimating fire ignition risk is to look at individual aspects related to human activities and vegetation characteristics separately.

In this case, the impact of urban development, road network density, and vegetation type were analysed individually.

7.2.2.2.1 Urban development:

According to a study conducted by KALABOKIDIS (2002) in Greece, urban areas can strongly influence fire ignition up to a distance of about 300 m, and beyond that distance, the impact decreases sharply.

EUFIRELAB

D-08-06.doc 51

We therefore quantified the number of fires occurring within that distance for the different types of urban setting shown in Figure 46.

For each urban setting, a coefficient was attributed according to the number of fires recorded within the 300 m limit; coefficient values ranged from -1 to 3.

The value of -1 was integrated was used in the very high density urban areas where the risk of a fire is considered zero, even though 2 fires were identified in that sector.

Table 10 summarizes the number of fires occurring within the 300 m limit of each urban type and for each fire ignition cause.

The bottom of Table 10 shows the coefficients that were attributed to these types

7.2.2.2.2 The road network.

The distance of 300 m from a house applied above in Table 10 was maintained for fires occurring in proximity to a road as shown in Table 11, which shows that there are far more fires occurring near municipal roads than any other type, and it is for this reason that they have been attributed a coefficient value of 3.

A coefficient of -1 was attributed to autoroutes because they tend to have large paved shoulders and few stopping areas in the Alpes-Maritimes, making them less likely to be sources of fire ignition.

Fire ignition was observed to occur near an autoroute at Aix-en-Provence,but this remains a rare event and was ignored in this study.

7.2.2.2.3 Vegetation risk.

As for the two preceding factors, a risk coefficient was attributed to the different types of vegetation according to the relative weight of fire ignitions in each type.

The number of fires and coefficients are described in Table 12.

The “moor” vegetation class had no impact on fire ignition and was eliminated from the analysis.

On the other hand, bushes (“maquis”) were strongly associated with fire ignitions and were attributed the greatest coefficient.

Shrub species favour fire ignition whereas tress species play an important role in fire propagation under certain vegetation humidity conditions.

7.2.2.2.4 The “structural” risk associated with fire ignition.

The structural risk map shown below (Figure 50) was elaborated working in raster mode with a cell size of 250 m, the resolution of the Corinne Land Cover.

The IGN BD Carto data base was in vector format and had to be rasterized.

Each raster cell was attributed the value of the coefficient attributed to it above.

The final ignition risk map was obtained by summing the individual coefficients of the variables defined above.

Summing the variables creates a raster layer where the cell values can range from -1 to 16 and where greater values indicate greater ignition risk: <0 = no risk, 0 = very low risk, 0≤4 = low risk, 4≤8 = intermediate risk, 8≤12 = high risk, ≥12 = very high risk.

7.2.2.2.5 Discussion

The fire ignition map in Figure 50 shows the greatest risk (orange and red) in the urban fringe of the coastal and pre-coastal areas.

The highly urbanised coastal areas are classed as no risk (green) due to the negative coefficient attributed to this sector where forest fires are extremely unlikely.

The fire ignition risk distribution can be analysed according to three types of human occupation.

In the South, the dense urban area reduces the fire ignition risk to zero.

In the extreme North, there is a forested zone that is too far from human activities to present much risk of a fire ignition that can be attributed to anthropogenic causes.

The low fire ignition risk is therefore associated with the low density of human activities.

Fire ignitions in this zone were due to lightning which strikes randomly.

Finally, there is an intermediate zone where fire ignition risk is at a maximum.

In this sector, human activities are sufficiently dense to create a high probability of fire ignition and yet low enough for the surface not to be largely non-vegetated and artificial, as in the urban centre.

There is sufficient vegetation to provide a serious fire risk and the human activities are neither dense enough to create largely artificial surfaces

Despite the logical distribution of forest fire ignition risk described above, the model has some faults.

Summing the coefficients of the three variables should give a maximum value of 9 (3 variables times a maximum coefficient value of 3), but the greatest value obtained is 16.

This indicates that some cells were counted more than once. In the dense road network of the urban fringe, cells located less 300 m from two or more roads would be counted once for each road.

The fire ignition risk is therefore overestimated in this sector.

In addition, the choice of a 300 m limit is somewhat arbitrary since this value, observed in Greece, may not be the most appropriate in the Alpes-Maritimes region.

Finally, the coefficient values varying from -1 to 3 were selected intuitively and may not reflect the actual weighted importance they represent.

7.2.2.3 The second modelling approach: statistical model A

After the initial intuitive approach, a statistical method was used to estimate fire ignition risk.

EUFIRELAB

D-08-06.doc 52

This second model is based on the number of fires occurring with varying distances of the two main factors, the road network and urban land-use.

In this approach, the number of fires occurring with successive bands of 25 m was calculated to estimate the range of impact of these factors on fire ignition.

The vegetation variable was not taken into account since its spatial distribution was not adapted to the approach.

In addition, no data were available on the density and height of the different vegetation layers (of the shrubs in particular) which could have been useful descriptors of the contribution of the vegetation to ignition risk.

7.2.2.3.1 Analysis of the factors at the departmental scale

For the entire Alpes-Maritimes department, the number of fires per band of 25 m from urban areas was counted (Figure 51).

Unlike the distance from urban settlements, Figure 52 shows a correlation between the number of fire ignitions and distance from the road network.

The number of fires decreases exponentially with increasing distance with a coefficient of determination (r2) greater than 0.95. Distance from roadways is therefore considered a good predictor of fire ignitions.

In order to improve the model, the Alpes-Maritimes department was separated into two sectors of different urban density.

7.2.2.3.2 Separating the department into two sectors.

Before separating the department into two zones, the dense urban coastal area was removed from the study since the risk of a forest fire ignition there is negligible.

The northern part of the department is mountainous and rural with low population and road densities as can be seen in Figures 53 and 54.

Between the northern part and the dense urban coastal area, an urban fringe can be found..

7.2.2.3.3 The urban fringe. - the impact of urban development : in this sector, the

number of fires decreases exponentially with the distance from human structures (Figure 55). The limits of the sectors appear justified since there is now a good relationship between the number of fire ignitions and distance from human settlements-

- the road network : fire ignitions decrease exponentially with distance from a road in the urban fringe.

7.2.2.3.4 The mountainous and rural northern sector of the department

- the impact of urban development: in the northern sector, there is no apparent relationship between the number of fire ignitions and distance from human settlements as can be seen in Figure 57. Human activities are therefore not a good predictor of fire ignitions in this area.

- the road network: as it was the case in the urban fringe, the number of fire ignitions decreases exponentially with distance from a roadway (Figure 58). Distance from the road network is therefore a good predictor of fire ignitions

For both the urban fringe and northern mountainous areas, the distance from a roadway is a good predictor of fire ignitions for a distance of up to 250 m from a road.

Urban development is a good predictor only in the urban fringe for a distance of up to 500 m.

In the northern sector, there is no apparent relationship between human settlements and fire ignitions, so this factor was eliminated in this sector in the following analyses.

7.2.2.3.5 The fire ignition risk map developed using statistical method A.

The fire ignition risk map using this method was elaborated in the following way.

A distance grid is created for each of the factors for the urban fringe and mountainous area.

It is important to note that these two geographical entities are treated separately.

Once that has been done, a new raster data layer is created by applying the regression equations shown above.

This results in three new raster layers: two for the urban fringe where both urban density and road network are significant factors, and one for the mountainous area where only road network is taken into consideration.

Each layer shows the potential number of fires estimated by the regression equations.

Classes of fire ignition risk are determined from the sum of the factors in the urban fringe and the fires predicted by the road network alone in the mountainous area.

The results of Model 2, method A are shown in Figures 59 (risk map alone) and 60 (risk map and actual fire ignitions)

7.2.2.3.6 Discussion.

The relationship between fire ignition risk and actual fires shown in Figure 60 appears satisfactory in the mountainous area since most of the fire ignitions occur with the high risk zone.

However, in the urban fringe several of the high risk zones have few or no actual fire ignitions.

The validity of the model in this sector is therefore put into question.

One of the reasons for the poor correspondence is perhaps because the road network and urban density are not independent since increasing urban density inevitably increases road density and vice versa.

This covariance is not taken into account in the urban fringe where the results of the predictive equations are summed.

In the mountainous zone, much of the surface area is in a low class risk, and some fires have occurred far from both houses and roads.

EUFIRELAB

D-08-06.doc 53

These can be attributed to either natural fires started by lightning or to humanly caused fires associated with pastoralism.

Shepherds sometimes ignite fires to maintain grass quality and occasionally these fires escape their control.

This hypothesis could be tested by determining the location of these fire ignitions with respect to the sheep grazing areas, but this information is not available through local agricultural authorities.

Since the statistical approach used in the second model proved deficient in the urban fringe, another statistical approach was tested.

7.2.3 The third model: statistical approach B.

This third model is also based on the distance from urban activities and roadways but it uses logistic regression which takes into account all of the cells.

This type of regression uses binary type data: absence or presence of a fire ignition between 1991 and 2003.

7.2.3.1 The method.

To optimize the calculations, the grid size was increased to 1 km2 squares.

The dependent variable in this regression is the number of fire ignitions between 1991 and 2003 in each 1 km2.

For each cell the mean distance from the centre of the cell to a road or settlement is calculated.

Table 13 shows that the tendencies observed above are still valid, but the relationship is weaker since cells without a fire ignition are also included in the regression.

In logistic regression, dependent variable is normally binary (0 or 1), but since some cells may have been affected by several fires (>1), the Y variable is transformed into a series of repetitions which each correspond to a Bernoulli law with the same P parameter, so that Y follows a binomial law with parameters n, and P, as shown in the equation below.

)2,1()2,1(1)2,1(

XXPXXPXXLogitP −

=

= Exp-(a0

+ a1

X1

+ a2

X2

)

The intensity of the risk is related to the

exp-(a0+a1*X1+a2*X2) term. 7.2.3.1.1 Results and risk maps.

The solution to the above equation gives the results summarized in Table 14.

The results of the logistic regression show that the values of the coefficients are statistically significant.

The values of their critical probabilities are always < 0.05.

In addition, the coefficients enable us to estimate the probability of fire ignition as a function of distance from urban activities or settlements (X1) and roadways (X2).

In order to obtain the erosion risk map, the equation is applied to the database.

We then obtain P values which range from 0.26 to 0.008.

These values are then subdivided into four fire ignition risk classes; added to these classes are the two classes which include the cells not taken into account (zero and very low).

Figures 61 and 62 show the spatial distribution of fire ignition risk based on this method..

Nearly half the surface area is in the low risk class, yet fires have occurred in these zones in the past. Most of these fires can be attributed to either lightning, which strikes randomly, or pastoralism, which tends to be located far from roads and houses.

The remaining area has a variable risk according the distance from roads and urban dwellings as predicted by the logistic regression.

Much of the urban fringe is in a high risk class, as it was for the first model.

Pressure from human activities in this zone makes it particularly vulnerable to fire ignitions.

The coefficient of determination (r2) for the logistic regression model is only 0.063, but this does not invalidate the model since the r2 value is considered less meaningful in logistic regression than in classical linear regression.

On the other hand, the probability associated with Pearson’s Khi2 value is pertinent and the risk of erroneously attributing explicative value to the road and urban variables is less than 0.1%.

7.2.4 Model comparison and conclusion.

From a simple visual evaluation of the risk maps, the three models appear to give similar results apart from differences in the spatial resolution.

In all three cases, the coastal area along the sea has no risk due to its high urban density, a mountainous zone in the North where the risk is low, and finally an urban fringe zone between the two others where the risk is particularly great.

Despite the similarities, some differences between the models persist.

The first two models have spatial resolutions of 250 m whereas the third has a resolution of 1 km.

The loss in spatial detail is compensated for by a more objective approach and valid statistical parameters.

Models 1 and 3 are the preferred options: the first for its simplicity and ease of application, the third for its rigour.

EUFIRELAB

D-08-06.doc 54

The three models agree with Figure 63 which shows the general risk behaviour associated with human pressure on the environment: in the first segment (mountainous), the risk is low due to low human pressure, the main cause of fire ignition.

In the third segment (coastal), the risk is also low, but in this case it is due to the scarcity of the vegetation in this densely urbanised environment.

In the second segment (urban fringe), the risk is at a maximum due to an abundant supply of combustible vegetation and the greater likelihood of fire ignition due to a high concentration of human activities.

Improving fire ignition risk modelling can only be done with better data acquisition techniques.

Both fire ignition location and cause must be investigated and recorded systematically in order to provide a fuller and more reliable data base.

In Spain, for example, forest fires causes are defined for the vast majority of fires, leaving very few unexplained.

Example of human ignition danger estimation at european scale

Continuing the work developed in Megafires Project (see section 2.1) a geo-statistical analysis to model wildland fire occurrence was carried out at provincial level in Southern Europe from socio-economic and demographic indicators of 1960 and 1990 together with variables that describe land cover and agricultural statistics.

This work is has been presented in Koutsias et al (2005) and is extended and inproved in a new paper actually in revision (KOUTSIAS et al. 2006).

Here is presented main results obtained in the first one.

In this study the classical OLS linear regression modeling has been used along with geographically weighted regression modeling to explain long-term wildland fire occurrence patterns at Southern Europe.

By applying the global OLS linear regression we identified critical underlying causal factors for wildland fire occurrence from a set of socio-economic, demographic, land cover and agricultural statistics.

Traditional regression modeling assumes that the relationship between the dependent variable and the explanatory variables is constant regardless of their geographical location.

This assumption, which is known as stationarity, is often violated in real world situations.

However, this constraint is overcome by another approach, known as geographically weighted regression (GWR), which considers that such relationships vary in space according to their location and that allows local variations to be taken into account.

In wildland fire occurrence modeling, especially when the geographical extent of the study area is large (i.e. the whole Southern Europe), it would be more reasonable to find varied rather than constant relationships.

In this study, non-stationarity is the main assumption made for modeling the driving factors behind wildland fire occurrence.

To overcome the constraints of traditional regression modeling which assumes stationary processes we applied geographically weighted regression (GWR) analysis.

GWR considers that relationships vary in space according to their location and allows local variations to be taken into account.

7.2.5 Study Area and Wildland Fire Database

The geographical extent of our study area includes Portugal, Spain, Southern France, South-Eastern Switzerland, Italy and Greece (see Figure 64).

Within the study area census data at provincial level were acquired from national statistics.

In total, 153 geographical units were identified, 18 in Portugal, 48 in Spain, 14 in South France, 20 in Italy, 2 in South-Eastern Switzerland and 51 in Greece.

For most of the countries these geographical units correspond to NUTS-3 level (Nomenclature of Territorial Units for Statistics) except for Italy where they correspond to NUTS-2 level and for Portugal where they correspond to their national district level.

Unfortunately, this not fully homogenized division among the countries, which resulted from limitations on data availability, could eventually influence negatively the regression analysis.

The basic data used in our study come from the MEGAFiReS project 1999) with an update for including the Southern part of Switzerland together with the number of wildland fires occurred within the period 1992-2000.

The database is composed by socio-economic and demographic indicators together with variables that describe land cover and agricultural activities.

In total, 77 variables were established for 1960 and 1990.

Differences of some variables (mainly for population characteristics) between the two date sets were also computed showing changes in population characteristics associated to urbanization and land abandonment.

These selected variables were considered of being potential underlying causal factors for explaining long-term fire occurrence patterns.

Besides the census data, the mean annual number of wildland fires, having occurred between 1992-2000 at provincial level, was also computed from national forest fires statistics.

This variable is used as depedent variable. For Portugal, Spain, Southern France and Italy

wildland fire observations were provided using the community centroids for the period 1992-2000.

For Switzerland the x and y coordinates of the ignition points were used having occurred in the same period.

Finally, for Greece, wildland fire observations were provided using the x and y coordinates of the ignition points for the period of 1985 to 1995.

EUFIRELAB

D-08-06.doc 55

Out of this original database, forest fire data were aggregated at provincial level so that to much the aggregation scheme used for the census statistics.

Wildland fire occurrence was expressed using the mean annual number of fires to overcome the inconsistence observed between the time period of Greek data and the rest of the countries.

A qualitative visualization of the spatial distibution of the dependente variable reveals critical regions for fire occurrence.

In future studies the mean annual number of fires is going to be estimated for two different fire size classes (burned area above 1 ha and above 100 ha).

7.2.6 Methodology: Geographically Weighted Regression (GWR), the linear and logistic case

GWR tries to capture the spatial variation by calibrating a multiple regression model so that at different points in space different relationships between variables can be found (ZHANG and SHI 2002).

A regression model is fitted at each data point by weighting all observations from that point as a fuction of distance.

Consequently, the neighbors sampled around the point influence the regression coefficients more strongly than the observations farther away.

GWR estimates the parameters at each point in the study area which then can be mapped using for instance geographic information systems (GIS) to investigate local spatial variation in the regression relationships (FOTHERINGHAM et al. 2002).

In geographically weighted regression the relationship between y and x can be expressed as:

where - β0(ui, νi), and βj(ui, νi) are estimated coefficients as

a function of location, and - ε is a random error term.

Besides the linear regression model we developed also a global and GWR logistic model.

In addition to linear regression, where the outcome variable is supposed to be continuous, logistic regression presupposes the existence of a binary or dichotomous dependent variable.

This main difference between the two regression models renders logistic regression very popular, especially when the experimental question can be expressed in a dichotomous way.

Besides this, the independent variables in logistic regression modelling can be a mixture of continuous and categorical variables.

Consequently, the assumption of multivariate normality is not presupposed in logistic regression.

Previous use of LR in fire occurrence at provincial level in Southern Europe can be found in Chuvieco et al. 1999b).

7.2.7 Results and discussion

7.2.7.1 Global versus GWR linear model

A stepwise ordinary least squares (OLS) regression was carried out for developing a model with the most significant variables.

Eight of the original variables have been chosen by the model which explains 53% of the variation of the dependent variable.

Among them the density of livestock (i.e., sheep) in 1990, the density of agricultural employees in 1990, the percentage of forested area, and the difference of the youth index between 1960 and 1990 are the most important variables based on the criterion of standardized coefficients of the model.

The GWR provided significantly better results than the global regression model, since the variance of the dependent variable explained increased to 68.65 %.

This is caused by the consideration of the non-constant relationships between the dependent and the explanatory variables, which change throughout the EUMed basin.

The residuals in GWR (Figure 64) are significantly less than those of OLS regression, indicating the better performance of the former model.

Based also on an ANOVA test, statistical significant improvement of the GWR model over the OLS linear regression model was found.

A Monte Carlo significance test for the parameters of the model found significant spatial variability on the intercept and on the coefficients of the difference for the youth index between 1960 and 1990, the number of agricultural employees in 1990, and the density of sheep in 1990.

Based on the same test, non-significant variability was found on the % of forested area, the % of the difference of the active population between 1960 and 1990, the density of agricultural employees in 1990, the % of difference in size of agriculture exploitations between 1960 and 1990, and the number of sheep in 1990.

7.2.7.2 Global versus GWR logistic model

Besides the OLS linear regression model we developed also a global and GWR logistic model. As mentioned, logistic regression presupposes a binary dependent variable that takes the value 1 in case of an event and 0 otherwise.

Thus we created a new variable which is based on the reclassification of the kernel density surfaces (two classes based on the equal area criterion for each country).

Then, the median value (1 or 0) found inside each polygon (= provinces) was attached as the dependent variable (Figure 65). We used the same independent variables as in the case of linear regression.

( ) ( )∑=

++=p

jiiijijiii uxuy

10 ,, ενβνβ

EUFIRELAB

D-08-06.doc 56

Three out of the four most significant variables of linear regression (the density of agricultural employees in 1990, the percentage of forested area, and the difference of the youth index between 1960 and 1990) were considered also by logistic regression as the most significant ones based on the Wald statistic. The GWR logistic model improved the deviance (-2LL) from 178.35 to 116.54.

Also, the GWR logistic model classified 84.97 % of the observations successfully, which is significantly higher than 66 % of the ordinary global logistic model.

The residuals in GWR logistic model, as shown in Figure 65, are less than those of ordinary logistic regression indicating the better performance of the former model.

7.2.7.3 Spatial Autocorrelation of Residuals

To further explore the advantages of the GWR over the classical global approach we calculated the correlograms of the residuals which show how they are distributed in space (Figure 66).

A reduction of Moran’s I spatial autocorrelation is evident in lag distances beyond 100 Km especially in the GWR linear model.

This reduction indicates that the model is appropriate.

The existence of spatial autocorrelation in OLS regression violates the assumption of independent errors (LICHSTEIN et al. 2002).

7.2.8 Conclusion

In conclusion, GWR seems to be a valuable approach for exploring and modeling non-stationary relationships between the response and explanatory variables and thus to better understand the spatial processes in wildland fire occurrence.

The reduction of Moran’s I spatial autocorrelation of the residuals indicates that GWR models were more appropriate than the classical global ones.

EUFIRELAB

D-08-06.doc 57

7.2.9 Figures

Figures 44 and 45: Topographic map (Source: BD Carto IGN) and Vegetation map (Source: Corinne LC)

Figure 46 and 47: Landuse types map (left) (Source Corinne LC) and road network map (right) of Alpes-Maritimes.

(Source: BD Carto IGN)

EUFIRELAB

D-08-06.doc 58

Figure 48: Recorded fires between 1991 and 2003 in the Alpes-Maritimes

Left: Figure 49: Spatial distribution of fires occurring during the 1991-2003 period according to cause

Right: Figure 50: Map showing the structural forest fire ignition risk The black points show actual fire ignitions, and the colours indicate the calculated risk

Figures 51 and 52: Number of fires versus distance to habitats (left) and to road network (right)

EUFIRELAB

D-08-06.doc 59

Figures 53 and 54: Urban density (left) and road network density (righ) in the Alpes Maritimes deparment

Figures 55 and 56: Number of fires versus distance from homes (left) and from a roadway (right) in the seashore

Figures 57 and 58: Number of fires versus distance to settlements (left), and to a roadway (right) in northern sector

EUFIRELAB

D-08-06.doc 60

R= ∑ F(Xi) ie Rlito = 45,545*Exp(-0,0126*X

1) + 23,65*Exp(-0,0053*X

3) (1)

Ra-pays = 38,10*Exp(-0,0101*X2

) (2) X

1 : Littoral road network X2 : Northern sector road network X3 : Littoral distance to habitations

Figures 59 and 60: Fire ignition risk map, model 2 (left) with 1991 - 2003 ignitions shown in white (right).

Figures 61 and 62: Fire ignition risk map obtained using the logistic regression model (left)

with fire departures between 1991-2003 (right)

EUFIRELAB

D-08-06.doc 61

7.2.10 Tables

Table 10: Number of fire ignitions according to urban density and ignition cause for fires less than 300 m from a residence.

Origin Dense urban Discontinuous urban Single dwellings Agricultural with

isolated dwellings Natural areas with isolated dwellings

Unknown 2 11 17 23 22

Voluntary 0 1 3 7 2

Other 1 0 2 10 6

BBQ 0 0 0 3 0

Gain 0 0 0 0 1

Electrical 0 0 0 0 0

Pastoralism 0 0 0 0 0

Lightning 0 0 0 0 1

Construction 3 0 4 12 9

% 5 10 18 34 32

Coefficients -1 1 2 3 3

Table 11: Number of fire ignitions according within 300 m of a road for each ignition cause.

Origin Autoroute National route Departmental Municipal

Unknown 3 13 47 112

Voluntary 0 2 4 21

Other 0 5 0 5

BBQ 0 6 10 21

Gain 0 0 2 0

Electrical 0 0 0 0

Pastoralism 0 0 1 15

Lightning 0 0 1 0

Construction 0 4 9 41

% 0.08 8 20 62

Coefficients -1 1 2 3

EUFIRELAB

D-08-06.doc 62

Table 12: Number of fire ignitions according to vegetation type and ignition cause.

Origin Coniferous Broadleaf Mixed Shrubs Grass Bush

Unknown 37 43 3 13 0 53

Voluntary 4 3 0 3 0 10

Other 1 1 0 0 0 0

BBQ 10 8 3 0 0 10

Gain 0 3 0 2 0 3

Electrical 1 0 0 0 0 1

Pastoralism 5 1 0 1 0 6

Lightning 5 5 0 0 0 1

Construction 11 8 1 1 0 7

% 27 27 3 7 0 35

Coefficients 2 2 1 1 0 3

Table 13: Correlation matrix

Fires Urban distance Road distance

Fires 1 -0,060 -0,109

Urban distance -0,060 1 0,119

Road distance -0,109 0,119 1 Bold font indicates a significant corrrelation at α=0.05 (bilateral test)

Table 14: Results of the logistic regression and equation giving the fire ignition probability as a function of two factors, X1 and X2.

Estimated model parameter values (maximum likelihood) :

Parameter Estimated value Standard deviation Khi² Pr. > Khi²

Constant -3,515 0,107 1073,719 < 0,0001

Urban distance (X1) 0,000 0,000 10,999 0,001

Road distance(X2) -0,005 0,001 50,738 < 0,0001

P(X1, X2) = 1 / (1+ Exp-(-3,515 - 0,000158*X1 - 0,005*X

2))

EUFIRELAB

D-08-06.doc 63

7.3 HUMAN IGNITION DANGER IN SOUTHERN EUROPE BASED ON FIRE OCCURRENCE MAPS

The objective of this and next Eufirelab deliverables is to elaborate a fire risk index at European scale.

In order to do it, and taking into account the resolution of satellite images and the size of the study area, it was decided that the ideal resolution of the pixel for the GIS layers must be 1 square kilometre.

In the case of the human ignition factors it is very difficult to get and have access to available and reliable data at this resolution or scale, for the whole or at least big part of the Euro-mediterranean area.

The main problem we find with the spatial variables (distances, fragmentation, interfaces) derived from cartographic layers like roads, urban areas, forest and land use covers, is that this original layer are very generalist at this global scale, with very poor detail. In the cases of socio-economic factors (unemployment, rural population age, population density…) estimated from census data, the main inconvenience is that they are only available in these moments for provinces.

This data unavailability at the desired scale (municipality level) for the whole area forced us to discard the socio-economic variables because we don’t consider appropriate to rasterize at 1 km2 the information derived from provinces.

The same problem regarding vector layers, like roads, with scales larger than 1: 1.000.000.

Because of these present difficulties to have an operational human danger ignition layer, one possibility can be generating it from historical data of fire occurrence.

From the location of ignition points and using the kernel density approach, that is a well known and explored interpolation method, we can generate a probability surface, as is showed in section 2.2 about Spread Project review.

Occurrence maps provides a global view of fires, which are mainly caused by human factors in Europe.

Visual exploration of spatial distribution of point observations is popular in point pattern analysis and takes place prior to any further analysis (BAILEY and GATRELL 1995, Fotheringham et al. 2000).

By visually examining point observations, one may acquire a prior idea about their spatial distribution, suggest hypotheses about explanatory factors, and decide for further statistical analysis and modeling (SADAHIRO 2003).

In the following section is explained deeply the generation of this map, according with the information of Deliverable D143 “Fire Risk and Human Factors (III)” from Spread Project.

7.3.1 Fire Occurrence Hot Spot Areas

For the analysis of large-scale fire occurrence patterns, we often need to transform fire ignition data that come in different formats and accuracies (e.g. x and y locations or number of fires per area unit) into continuous surfaces.

As a non-parametric smoothing technique, the kernel density estimation (SILVERMAN 1986) is well suited for converting fire ignition point data into continuous “fire occurrence distribution surfaces”.

Kernel density estimation is based on the estimation of the density at each intersection of a grid superimposed on the data, after placing a probability density (kernel) over each point event.

Depending on whether constant or multiple adaptive values are used for the smoothing parameter, kernel density estimation is distinguished into the fixed and adaptive method, respectively. (SILVERMAN 1986, Levine 2002).

KOUTSIAS et al. (2004) applied the kernel approach to assess fire occurrence patterns at landscape level by addressing some of the inherent positional inaccuracies of the fire ignition locations.

DE LA RIVA et al. (2004) used also this approach to express fire occurrence patterns at municipality level by using fire ignition observations.

The current study proposes an adaptive kernel density interpolation approach applied to community centroids where the number of fires per community is used as the intensity variable for the kernel density estimation.

The final aim is to define large-scale fire occurrence patterns and to identify ‘fire occurrence hot spot areas’ in southern Europe, using fire ignition observations aggregated at community level. In our study the number of fires form the period 1992-2000 have been estimated from national fire statistics and expressed at community level using the community centroids (Figure 67).

The geographical extent of the study covers the European Mediterranean countries including Portugal, Spain, Southern France, Italy, Greece, as well as South-Eastern Switzerland (Canton of Grison and Ticino) from central Europe (Figure 67).

Motivated by the non-homogeneous spatial distribution of community centroids (Figure 67) we decided to choose the adaptive kernel density estimation mode instead of the fixed one.

The adaptive mode allows for the adjustment of the bandwidth size in relation to the concentration of the community centroids (WORTON 1989).

For the adaptive kernel approach different bandwidth sizes have been tested (from 1 to 40 at various steps) that are indicative of typical resulting patterns of under- to over-smoothing.

Here we present four representative steps (2, 5, 10, 20).

For a homogeneous process with no spatial dependence, the expected number of events within a distance d of a randomly chosen event equals ?pd2 where ? is the density (GATRELL et al. 1996).

Thus, if the expected number of events corresponds to the number of centroids used in the adaptive kernel, the estimated distance d can be used for defining the bandwidth size in the fixed mode.

The corresponding bandwidth sizes in the fixed mode are 3622, 5726, 8099 and 11453.

EUFIRELAB

D-08-06.doc 64

We applied both approaches, the fixed and the adaptive kernel density interpolation, at national level in Portugal in order to decide about the most proper approach and the most appropriate bandwidth size (Figure 68).

The effect of small bandwidth size to kernel density surfaces is evident in the corresponding images in Figure 68.

Low bandwidth sizes for both, the adaptive and the fixed kernel density, (e.g. comprising 2 and 5 community centroids or 3622 and 6726 meters), created surfaces in which the local variation of the density surfaces is enhanced around the interpolated points.

This effect is more evident in the fixed kernel density approach because the bandwidth size remains constant throughout the whole extent of the study area while the concentration of the interpolated points (community centroids) varies significantly.

This variation in the concentration of community centroids is taken into account by the adaptive kernel mode since the bandwidth size varies locally so each time the same number of community centroids to be used.

Based on visual inspection, among the different bandwidth sizes used the one that corresponds to 10 community centroids is considered of being the most appropriate to avoid under- or over smoothing.

To evaluate the density surfaces three grids of 2, 5 and 10 Km resolution were overlaid over the density surfaces.

Subsequently the number of fires and the sum of density kernel values were estimated.

The correlation coefficients (Table 15) were slightly higher between the number of fires and the adaptive kernel density surfaces.

This is another indication that the adaptive kernel density interpolation performs better compared to the fixed kernel density as the former allows for locally varying bandwidth sizes.

The same methodology described previously and applied at national level in Portugal has been applied to the rest of the countries.

The absolute values of kernel densities for each country depend on the number of fire ignition points as well as on the spatial distribution of community centroids and wildland fire observations.

The total number of wildland fire ignition observations varies greatly between the different countries and consequently kernel density values are also very diverse.

The consequence of this variability found among the countries results in an underestimation when the kernel density surfaces are joined together under a common classification scheme as shown in Figure 69.

This problem can be overcome if a reclassification of kernel density surfaces is applied before merging all data.

Thus kernel densities have been reclassified to 10 classes based on the equal area criterion within each country (Figure 69), presupposing equivalence for fire hot spot areas among the countries.

EUFIRELAB

D-08-06.doc 65

7.3.2 Figures

Figure 63: Fire ignition risk evolution according to vegetaion and human activity characteristics. (Carrega, 1992)

Figure 64: Observed and predicted values of OLS and GWR linear models and the distribution of their residuals.

Figure 65: Observed and predicted values of ordinary and GWR logistic models and their residual distribution.

EUFIRELAB

D-08-06.doc 66

Figure 66: Correlograms of the residuals

Figure 67: Community Centroid

Figure 68: Kernel density estimates using the fixed (upper images) and

the adaptive kernel approach (lower images).

EUFIRELAB

D-08-06.doc 67

Figure 69: Large-scale fire occurrence patterns at European level by kernel density estimation

7.3.3 Tables

Table 15: Correlation coefficients between number of fires and kernel densities at 2, 5 and 10 km grid resolutions

EUFIRELAB

D-08-06.doc 68

8 PROPAGATION DANGER INDEX

The propagation danger component hast to cover the potential hazard that a fire propagates in space and time.

This index is associated to estimating the behaviour of the fire.

Fire behaviour is mainly described with reference to the propagation rate and the intensity of the flame front.

Not too many studies include this type of index. At global scale, a case applied to Euro-Mediterranean basin has been included in the Spread project, described in chapter 4.

The propagation danger index (PD) is composed by two subindex: the average rate of spread (RoS), and the flame length (FL).

Both variables are calculated using BEHAVE algorithm.

The input data is obtained from CORINE and a global Digital Terrain Model.

The results is a map taken as static, it will not change throughout the fire season.

Taking into account the characteristics of the index described in Spread project (input variables, spatial and temporal resolutions), we propose this index as the “Propagation Danger Index” ot he Euro-Mediterranean Wildland Fire Risk Index (EM-WFRI) defined in this deliverable.

The definition and estimation of this index are also described in the same chapter 4.

8.1 AVERAGE RATE OF SPREAD AND FLAME LENGTH

In order to obtain a global view of risk associated to fuel loads, terrain characteristics and wind flows, a global simulation analysis was performed.

This analysis tried to obtain average values of rate of spread and flame length, considering different wind and topographic conditions for the estimated fuel maps of the whole EUMed area.

This attempt should be considered as a general overview of average expected fire behaviour at global scale, in order to rank different danger levels according to the combination of fuel and terrain spatial patterns.

The estimation of the average RoS was based on several simulations performed by the Autonomous University of Barcelona for different fuel types, slope ranges and wind flows.

As a simulation kernel the wildland simulator proposed by COLLINS D. BEVINS, which is based on the fireLib library (COLLINS, 1996) were used.

FireLib is a library that encapsulates the BEHAVE fire behaviour algorithm (MORGAN et al., 2001).

In particular, this simulator uses a cell automata approach to evaluate fire spread.

The terrain is divided into square cells and a neighbourhood relationship is used to evaluate whether a cell will be burnt and at what time the fire will reach the burnt cells.

As inputs, this simulator accepts maps of the terrain, vegetation characteristics, wind and the initial ignition map.

The output generated by the simulator consists of two maps of the terrain.

In the first one, each cell is labelled with its ignition time; in the second one, each cell is labelled with its flame lenght.

This information must be used to calculate the rate of spread and an average from among all flame lenght.

To calculate the rate of spread, the distance between the ignition point and each particular cell in the terrain is divided by the ignition time of that particular cell.

This calculation is repeated for each cell in the terrain to determine the maximum value of the rate of spread.

This maximum value is used as the rate of spread for that particular situation.

To provide the propagation danger map, a set of prototype plots was created, considering all the fuel models from Rothermel classification and a certain slope percentage (from 0 to 100%, with a step of 5%).

The total number of plots was 273. Each plot consists of a grid of cells with 11 columns x 11 rows (each cell measured 328.083 x 328.083 feet).

The ignition point was located in the middle of the plot.

EUFIRELAB

D-08-06.doc 69

For each plot, many input parameter combinations were used to simulate the wildland fire behaviour and the average rate of spread and flame lenght were also calculated.

The parameters considered for variation were: 1-hr dead fuel moisture, 10-hr dead fuel moisture, 100-hr dead fuel moisture, live herbaceous moisture and wind speed and direction.

Average values of rate of spread (RoS, in m/min) were computed for the different fuel types.

Finally, the RoS values were scaled into a 0-1 range, by normalizing the values between the maximum and minimum values (figure 70).

Fuel types were derived from the Corine land cover. The resulting map is considered static, since no

specific conditions are simulated (wind or FMC), but only general patterns of propagation rates.

Same process as above was performed to compute the average flame length (FL), measured in metres and normalized into a 0-1 scale (figure 71).

8.2 PROPAGATION DANGER (PD)

It derives from the combination of the two intermediate products previously described: RoS and FL.

The results of the simulations were mapped at EUmed scale using CORINE land cover (reclassified into fuel models) and slope maps

The maps of FL and RoS were then normalized using linear fitting and multiplied to produce PD:

PD = [(RoSi – RoSmin)/(RoSmax - RoSmin)+0.001] * [(FLi – FLmin)/(FLmax - FLmin)+0.001]

A small constant (0.001) was added to avoid zero multiplication in case of minimum values.

RoS and FL were considered in this formula of equal importance, although this could be tuned up in future improvements, according to further experience or suggestions.

This map is taken as static, i.e. it will not change throughout the fire season.

8.3 FIGURES

Figure 70:Estimated average Rate of Spread (Normalised values from 0 to 1)

Figure 71: Estimated average Flame Length (Normalised values from 0 to 1)

EUFIRELAB

D-08-06.doc 70

9 VULNERABILITY INDEX

9.1 POPULATION VULNERABILITY

9.1.1 Introduction

A vulnerability analysis is realized to obtain information about the susceptibility of individuals, property, and the environment to the adverse effects of a given hazard in order to develop appropriate prevention strategies.

The analysis of this information helps to determine who is most likely to be affected, what is most likely to be destroyed or damaged, and what capacities exist to cope with the generated effects.

To perform a proper and detailed vulnerability analysis it should be necessary: - to identify that part of population which is

vulnerable because of proximity or exposure factors;

- to assess vulnerability of buildings, infrastructure and other aspects of the physical environment due to factors such as site, materials used, construction technique, and maintenance;

- to evaluate transportation systems, communications systems, public utilities (water, sewage, power), and critical facilities (i.e., hospitals) for weaknesses; to identify most vulnerable population (e.g., older adults, children, single-parent families, the economically disadvantaged, the disabled);

- to estimate the level of poverty, jobs that are at risk, and the availability of local institutions that may provide social support;

- to determine the community’s potential for economic loss and recovery following a disaster;

- to identify existing measures and resources that can reduce the impact of a given hazard.

The degree to which populations are vulnerable to hazards is not solely dependent upon proximity to the source of the threat or the physical nature of the hazard, social factors also play a significant role in determining vulnerability (CUTTER et al., 2000).

The primary social effects of wildfires are physical: houses and other kinds of human structures will burn; people may be injured or may die.

However, it should be noted that death and injury depend upon unique sets of circumstances surrounding each fire event: these cannot be modelled here.

For instance, as noted by CASE et al. (2000), in a geographic information system, we can only describe the total number of people at potential risk, and the number of people having special characteristics which place them in jeopardy.

Some individuals belong to special populations because they have characteristics, which lead them to suffer effects disproportionate to the general population.

Special populations include families with very young children, the elderly, and households whose members have incomes below the poverty line.

Risks to these populations, as pointed out by CASE et al. (2000), is expressed as a greater difficulty of successful evacuation, greater susceptibility to health impairment, loss of possessions (which could be irreplaceable for people in two of these special groups), and loss of jobs and income.

In brief, houses which are dependent upon domestic water supplies and fuel-wood, which lie in remote areas, and which have difficult routes of ingress and egress are at greater risk of damage or destruction in the event of wildfire.

People who are very young or very old are more likely to experience adverse health consequences in the event of wildfire, and people who have very low incomes are more likely to suffer irreversible economic and social consequences. When these groups of people suffer from wildfires, they have less individual means of recovery, and require more help from the communities they live within (CASE et al., 2000).

To identify those sections of a community most likely to be affected by a particular hazard and to determine areas that require strengthening to prevent or mitigate the effects of the hazard, data on several variables must be collected (among which, information on the size, density, location, and socio-economic status of the at risk community).

Vulnerability assessment is one of the least investigated tasks, due partly to the lack of relevant detailed socio-economic data and the difficulty of their effective spatial representation for integration with physical environmental data on hazards (CHEN et al., 2003).

Vulnerability assessment is however an important task in risk assessment and has social significance for a hazard-prone vulnerable community.

Assessing vulnerability in spatial terms requires a wide range of physical and socio-economic knowledge and expertise.

In this context, GIS can be used for database establishment, analytical modelling, and decision support in a decision-making process.

GIS can play an important and integral role in lessening the adverse impacts of natural hazards on society.

Having a wide range of spatial analysis techniques and tools, they are helpful for identifying, measuring, and assessing many aspects of natural hazards and their consequences (CHEN et al., 2003).

GIS spatial analysis in particular, applying various methods and techniques, has the ability to employ physical environmental and socio-economic data for risk and vulnerability analysis.

High-quality GIS databases support subsequent risk assessment and rational decision-making in a spatial and temporal context, which can help risk managers and the public understand how complex hazards and their consequences will affect vulnerable communities.

EUFIRELAB

D-08-06.doc 71

Spatial information system can be used to estimate the physical risk to people and houses, and to draw useful inferences about the nature of secondary, less tangible, and less predictable social effects resulting from wildfire (CASE et al., 2000).

To reach these tasks geo-referenced data layers must be collected from a variety of sources.

When we have to assess vulnerability to wildfire, and in general in all the natural hazards context, the required data can be divided into three categories - physical environmental data, socio-economic environmental data, and management-related data (CHEN et al., 2003).

Socio-economic environmental data are in particular used to assess community vulnerability and include population and housing census data and data on utilities and access.

Because of the increasing emphasis on community vulnerability assessment, collecting socio-economic data is essential.

According to Granger 1998) detailed information on setting, shelter, sustenance, security, and society is required.

For example, data on shelter include construction materials of the walls, roofs, and floors, and the ages of buildings.

Data on utilities (e.g., water, electricity, telecommunication, gas), data on security facilities (e.g., hospitals, police stations, emergency operations centres), and data on access (e.g., roads, bridges, tunnels, railways) are also required, wherever possible.

The concepts previously expressed can be usually applicable to a local scale, as a consequence of the great amount of required input data and their very refined spatial resolution.

The possibility to obtain a refined result is typically dependent on the chosen working scale and is strictly related to the availability and the resolution of data.

The scale and resolution of the data that can be used (typically available at the state scale) limit the detail that can be produced.

This choice can be conflicting with the desire of scientists to be precise, and the needs of land managers to identify specific project activities and locations (SAMPSON and NEUENSCHWANDER, 2000), but is consistent with the objective of producing a strategic assessment, at least concerning population vulnerability.

At the 1:1,000,000 scale, corresponding to the one selected to develop the Euro-Mediterranean Wildland Fire Danger Rating System, the social data sets above all are necessarily not detailed enough to allow us to quantify the complete suite of social effects we know can occur.

Additionally, some crucial social effects depend very much on the exact nature of the circumstances surrounding each fire or upon problematic factors that cannot be adequately described in any known model (CASE et al., 2000).

Some general and basic census data can be utilized to develop indicators of relative risk or vulnerability to human populations.

The elements chosen for this analysis can be the total population and population density within each unit (for instance a 1 km2 grid) and populations that might be placed at additional risk due to their circumstances (children, elderly, low-income).

These factors can be chosen as examples, not as an exhaustive listing of relevant social information; however they can give some information about the ways in which people experience wildfire, and how they may be affected by it, providing decision makers with useful information in making judgments about prioritising mitigation and protection programs (CASE et al., 2000).

We could construct a density map of these special populations, reasoning that proportionately greater risk occurs where people and houses with these particular characteristics occur in greater numbers (CASE et al., 2000).

Another crucial data set is represented by the number of ignitions per 1 km2 grid over a certain historical period.

On the assumption that areas experiencing the most ignitions in the past will continue to experience the highest relative ignition rates, areas most likely to experience future wildfire ignitions can be identified (NEUENSCHWANDER et al. 2000).

The importance of assessing population vulnerability is particularly relevant when we have to face the problem of wildland-urban interface (WUI) areas.

Wildland-urban interface in the European environment and more specifically in the Mediterranean area is a very complex spatial context with many interrelated social, natural resource and wildfire issues.

The problem of wildfires in the wildland-urban interface has recently become quite relevant because of the increasing number of dwellings near to or inside natural areas and of the increasing number of wildfires involving these sites.

As previously underlined, Geographic Information Systems (GIS) can be useful tools for WUI management, through their capability of handling in an integrated environment multi source and multi resolution spatial data (BURROUGH and MCDONNEL, 1998).

EUFIRELAB

D-08-06.doc 72

9.1.2 Example of population vulnerability mapping in Piemonte Region (Italy)

As an example of GIS analyses that can be performed in WUI areas to assess wildfire risk and vulnerability see CAMIA et al. (2004).

The analyses described in this work were performed in the framework of WARM Project.

They were firstly aimed at identifying and typifying settlements, providing a first description of the surrounding environment and of fire suppression preparedness.

As previously mentioned, the study area is characterized by small urban settlements, presenting both situations of classical interface and intermixes.

Actually, an entire set of conditions, ranging from the densely inhabited village with boundaries facing wildland and forests until the isolated houses are represented.

The authors wanted to identify and classify the different situations according to the degree of urbanization within the wildland area, and therefore the general wildland-urban mixing conditions for the different settlements, being clear the usefulness of such a characterization from a fire management point of view.

This task was approached using density analysis techniques to the house layer of the database, and applying thresholds to the derived house density layer and identify in this way settlement boundaries.

Settlement types were then characterized by computing the number of houses in each polygon discriminated, performing a spatial join between the settlements and the houses layers.

The presence of specific fuel types, as well as the topographic arrangement of the area, provides useful indications about the fire hazard and risk conditions around the settlements, but also about the possibility and strategies for fire fighting resources.

In addition to such key features, from a fire suppression point of view, it is important to know the accessibility of the settlements to be protected, as well as the relative location of fighting resources (fire stations) and water supplies.

Although a road density map can provide useful information on the accessibility of the sites, much more informative is the map derived from the network analysis that allows to classify each settlement according to the time needed to reach it from the nearest fire station or the time from the nearest water supply.

Working at a global scale within a 1-km² grid, an example of what can be obtained with simple and necessarily coarse data, is reported in Camia et al. (2002).

In this work, a methodology was developed to analyse the problem of WUI wildfires and to characterize its spatial distribution at the scale of the Italian Regions.

Italy is divided into 20 Regions, characterized by deep environmental differences; the study was applied in Piemonte, a Region located in the north-western part of Italy, with a surface of about 2,5 million hectares.

It has a quite heterogeneous territory, characterized by plains mainly interested by agriculture, hilly areas and mountains among the highest of Europe.

There are 1.209 municipalities, whose population ranges from about 1.000.000 inhabitants to less than 100 inhabitants.

In the country areas the secondary houses are quite frequent and their number has been strongly increasing in the last years.

The fire season is typically a winter-early spring one, which is mainly characterized by surface fires.

Crown fires are mostly due to the presence of evergreen coniferous stands, while most forests are made up by broadleaved species.

Most wildfires break out and spread in the mountain belt from 200 m to 1000 m a.s.l..

Data about wildfires are collected and recorded with a spatial resolution of 1-km2 (i.e. on a UTM kilometric grid with square cells of 1000 m size).

For each wildfire, the cell where the fire started is recorded, while the adjacent cells that could have been interested by the fire spread are not reported.

Only recently, information on fire perimeter was introduced among data collected following a fire event.

Among the available data there is no information concerning possible infrastructures damaged or threatened by fire.

WUI wildfires are therefore not identifiable from historical series.

Considering these constraints due to data deficiency or resolution, a methodology that analyses past events examining and assessing their territorial background was proposed.

The geographical units of the analysis correspond necessarily to 1-km² cells.

To locate WUI fire prone areas in Piemonte territorial elements that contribute to create the WUI environment were identified.

Areas where natural vegetation and infrastructures are both present and intermingle were looked for using and overlaying layers containing urban areas and woods, with predefined criteria.

Thus, working mainly with GIS facilities, proper layers derived from wildfires, urban areas and forest databases were overlaid applying thresholds and neighboring constraints in order to extract those cells that are expected to contain WUI environment prone to wildfires.

Data processing that eventually led to the selection of territorial contexts where WUI areas are potentially threatened by wildfires, allowed to assess the size of the problem at regional scale, its spatial distribution and to find the most affected areas.

EUFIRELAB

D-08-06.doc 73

The first GIS layer analysed is represented by wildfires data.

An historical series of 20 years, from 1980 to 1999, was considered.

From these data the number of wildfires occurred in this period within each grid cell was obtained.

With this respect, it is important to underline once again that it is possible to know only the cell in which the wildfire started and not the neighboring cells involved, if any.

To include also those cells that, although not directly recorded as sites where fire actually started, were expected to be affected by fires started in neighbouring cells, some further criteria were introduced.

The second input layer is represented by the territory urbanization.

Urban areas have been extracted from the land use map (1:100,000) of Piemonte Region cartography.

These areas were intersected with the cells containing wildfires in order to select those cells characterized both by houses and by wildfires.

In this step all cells containing exclusively urban area and not burned wildland were excluded.

In this way, starting from all wildfires occurred in Piemonte from 1980 to 1999, cells with wildfires occurred next to urbanized areas were identified.

As mentioned, wildfire starting within cells with no urban area but adjacent to cells with urban area presents a potential menace.

Therefore, we added to the previous selection, cells less than 1,000 m away from other cells characterized by the presence of urban areas and with more than 3 wildfires or more than 10 ha of burned surface in the 20 considered years.

This step allows to include not only directly affected areas, but also those urban areas considered potentially prone to WUI wildfires.

The third input layer is related with forests. The forested surface of each cell was calculated

from the land use map, selecting between the previously extracted cells only those typified by a forestry cover higher than 5%.

Considering starting data and analysis scale, it isn’t possible to assert these wildfires directly interfered with human dwellings, but just that they occurred in environments characterized by the presence of WUI areas.

To characterize areas where WUI wildfires are particularly critical, all cells between those previously selected, in which events characterized by a burned surface higher than 30 ha and 15 ha occurred, have been identified.

These two values of burned surface were chosen since 30 ha correspond to the burned surface threshold that typifies a critical fire according to the Regional Fire Management Plan (BOVIO et al., 1999), while 15 ha have been prudentially added as an in-between attention threshold for WUI fires in Piemonte.

On the base of WUI wildfires frequency, a zoning of municipalities has been realized.

Piemonte municipalities were classified, identifying 5 frequency classes, according to the number of wildfires potentially interesting WUI areas between 1980 and 1999.

9.1.3 Estimation of population vulnerability at Euro-Mediterranean scale

Vulnerability is not directly measurable. However, it is possible to induce vulnerability using

information about the characteristics of the geographical zone considered.

As previously underlined, to realize a Wildland Fire Danger Rating System operative at an Euro-Mediterranean scale, we must face the problem of using appropriate data at relevant scale and precision for the expected aim, currently available for all the countries involved.

Thus the question is: What do we have and how reliable is it?

The difficulty when trying to approach the risk at global scale is to find relevant and available indicators allowing a comparison between all countries.

Simple indices relying on good data and with stated limitation and subjectivity might be much more efficient than complex ones that cannot be computed because of the lack of (reliable) data.

Extrapolations from local researches to global scale are rarely applicable as data may not be of comparable formats or simply not available.

If a model requests a large amount of inputs, the chances that such model will never be used, by lack of data or by too fuzzy data, are significant.

On the other hand, a model based on too few parameters will lead to large gap between observed facts and expected figures.

Since no more detailed and precise data are currently available and/or easily accessible for all the Euro-Mediterranean countries, it was decided to work with the CLC2000 geographic data layer (EC JRC-IES, 2005).

CORINE land cover (CLC) is a geographic land cover/land use database encompassing most of the countries of the European Community and the majority of the Central and East European countries (Figure 72).

CLC 2000 is the year 2000 update of the first CLC database, which was finalised, in the early 1990s as part of the European Commission programme to COoRdinate INformation on the Environment (Corine).

It also provides consistent information on land cover changes during the past decade across Europe.

The CLC2000 database currently covers 32 countries.

EUFIRELAB

D-08-06.doc 74

CLC describes land cover (and partly land use) according to a nomenclature of 44 classes organized hierarchically in 3 different levels (Figure 73).

The first level (5 classes) corresponds to the main categories of the land cover/land use (artificial areas, agricultural land, forests and semi-natural areas, wetlands, water surfaces).

The second level (15 classes) covers physical and physiognomic entities at a higher level of detail (e.g. urban zones, forests, lakes), while level 3 is characterized by 44 more detailed classes.

The nomenclature has been developed in order to map the whole Community territory, including the foreseen extension to other eligible countries.

The use of the CLC nomenclature with 44 classes at three hierarchical levels is mandatory.

Additional national levels can be mapped but should be aggregated to level 3 for the European data integration.

No unclassified areas should appear in the final version of the data set.

CLC was elaborated based on the visual interpretation of satellite images.

Ancillary data (aerial photographs, topographic and vegetation maps, statistics, local knowledge) were used to refine interpretation and the assignment of the territory into the categories of the CORINE land cover nomenclature.

The smallest surfaces mapped (mapping units) correspond to 25 hectares.

Only area elements (polygons) are identified. Areas smaller than 25 ha are allowed in the

national land cover database as additional thematic layers, but should be aggregated/generalized in the European database.

Linear features less than 100 m in width are not considered

The scale of the output product was fixed at 1:100.000.

Thus the location precision of the CLC database is 100 m.

From the CLC2000 layer, information on both potentially combustible land covers and populated ones could be derived, providing a common base for all the involved countries, from which to indirectly obtain information on population vulnerability to wildfires.

An index to assess population vulnerability was defined, made up by three main sub-indices.

These three sub-indices are calculated on a 1-km2 grid and their values within each cell are summarized to give the final index.

9.1.3.1 SUB-INDEX 1

This sub-index qualifies each grid cell according to the presence within the cell itself of urban land covers and potentially combustible ones. The presence of these land covers is quantified with this sub-index independently from their spatial location and topology within the cell.

Third level CLC2000 classes were used.

The surface occupied inside the cell by the following land cover classes is calculated in a GIS environment: - Areas occupied by forests and semi-natural areas,

shrubs and herbaceous plants; recently burned areas are also included (CLC2000 codes: 2.4.4; 3.1.x; 3.2.2, 3.2.3, 3.2.4, 3.3.4 ; see Table 16);

- Areas characterized by a major presence of houses and consequently population (CLC2000 codes: 1.1.x; 1.2.x; see Table 16);

This sub-index is meant to take into account those situations characterized by a significant presence of both combustible vegetation and population potentially at risk.

The total surface occupied within each cell by the two land cover categories (a and b) is then classified in ten percentage classes.

The value of the sub-index, going from 0 to 5, is obtained through a bi-dimensional matrix (Figure 74), combining the classes of the two land cover categories present in a cell.

In the matrix higher values are assigned to those cells in which there is a major presence of both urban land covers and potentially combustible ones, in order to identify those situations characterized by a higher probability to have wildland-urban interface areas or by a higher probability to have a widespread human presence, independently from the precise location of settlements.

9.1.3.2 SUB-INDEX 2

This sub-index is based on the assessment of the spatial configuration of specific land covers in the cell.

The linear development of those zones in which urban land covers are in contact with potentially combustible land covers is measured.

Considering the scale and precision of CLC2000 data, this contact is more probably to be considered as a neighbourhood relationship between the two categories of land cover.

The land covers selected for this analysis are: - Areas occupied by forests and semi-natural areas,

shrubs and herbaceous plants; recently burned areas are also included (CLC2000 codes: 2.4.4; 3.1.x; 3.2.2, 3.2.3, 3.2.4, 3.3.4 ; see Table 16);

- Areas characterized by urban fabric, continuous or discontinuous (CLC2000 codes: 1.1.x; see Table 16);

The main aim of this sub-index is to identify those situations in which, given a certain surface occupied within a cell by the two categories of urban and potentially combustible land uses, as assessed by sub-index 1, there is a high probability that people, buildings or infrastructure are in proximity of fuels susceptible to burn.

EUFIRELAB

D-08-06.doc 75

Please note that the susceptibility and fire-prone nature of these fuels should be assessed though other components of the Euro-Mediterranean Wildland Fire Risk Index proposed in this deliverable.

As a consequence, the effective population exposition to fire risk comes from the integration of the Vulnerability index here discussed and the Ignition and Propagation Danger indices previously discussed.

The sub-index value is assigned according to the linear development of probable WUI areas measured in the cells and classified following Table 17.

Only those cells having at least one of the two above described sub-indices different from 0 are selected to proceed with the analyses.

This means that cells with no or little presence of urban and potentially combustible land uses and cells with no or little linear development of WUI areas are considered as not vulnerable areas, given the low detail of initial data and the working scale.

9.1.3.3 SUB-INDEX 3

Cells with a certain level of possible vulnerability, as resulting from the previous steps, are then further differentiated according to the presence of a particular urban land cover, i.e. the “discontinuous urban fabric” (CLC2000 code 1.1.2). This land cover identifies areas where most of the land (between 30 to 80 %) is covered by structures; building, roads and artificially surfaced areas are however associated with vegetated areas and bare soil, which occupy discontinuous but significant surfaces. This land cover, frequently corresponding to the external part of settlements, is considered a more vulnerable urban cover because of the concurrent presence of structures and artificial green areas or wildlands, often associated in a complex spatial pattern.

The surface occupied by discontinuous urban fabric inside cells previously selected is then calculated, assigning to each cell a value corresponding to a surface class as defined in Table 18.

The final vulnerability index was obtained as the sum of the three sub-indices in the cell, ranging from a minimum of 1 to a maximum of 14.

Three vulnerability classes were identified, as reported in Table 19.

An example of each sub-index and of the final population vulnerability index, calculated according to the methodology above described, is reported in Figure 75.

The described methodology was applied to the Italian province of Torino (North-Western Italy, Piemonte Region), corresponding to a NUTS 3 level. The CLC2000 data layer of the province and the DTM are shown in Figure 76.

Resulting values of sub-index 1, sub-index 2, sub-index 3 and the final vulnerability index are instead reported in Figure 77.

EUFIRELAB

D-08-06.doc 76

9.1.4 Figures

Figure 72: Example of CLC2000 geographic data layer. Yellow lines correspond to the boundaries of CLC classes

(see figure 73 for CLC nomenclature); red lines correspond to 1-km2 grid cells.

Figure 73: CLC2000 nomenclature (from EC JRC-IES, 2005)

EUFIRELAB

D-08-06.doc 77

1 2 3 4 5 6 7 8 9 10

surface %

0 - 1

0

11 -

20

21 -

30

31 -

40

41 -

50

51 -

60

61 -

70

71 -

80

81 -

90

91 -

100

1 0 - 10 0 0 0 0 0 0 0 0 0 0

2 11 - 20 0 1 1 1 2 2 2 1 1

3 21 - 30 0 1 2 2 2 3 3 3

4 31 - 40 0 2 2 3 3 4 4

5 41 - 50 0 2 2 3 4 5

6 51 - 60 0 2 3 4 5

7 61 - 70 0 2 3 4

8 71 - 80 0 3 49 81 - 90 0 3

10 91 - 100 0

urban areas

com

bust

ible

are

as

Figure 74: Bi-dimensional matrix to determine sub-index 1 values within each 1-km² grid cell.

Figure 75: Example of each sub-index and of the final population vulnerability index calculated according to the

methodology here described

EUFIRELAB

D-08-06.doc 78

Figure 76: DTM and CLC2000 data layer of the Italian province of Torino (North-Western Italy), where the

population vulnerability index here proposed was tested

EUFIRELAB

D-08-06.doc 79

Figure 23: Sub-index 1, sub-index 2, sub-index 3 and the final population vulnerability index computed for the

province of Torino (Italy)

EUFIRELAB

D-08-06.doc 80

9.1.5 Tables

Table 16a: Corine land cover nomenclature of the classes selected for the analyses (Populated Land Covers)

CLC code CLC description

Areas mainly occupied by dwellings and buildings used by administrative/public utilities or authorities, including their connected areas (associated lands, approach road network, parking-lots). 1.1.1 Continuous urban fabric: Most of the land is covered by structures and the transport network. Building, roads and artificially surfaced areas cover more than 80 % of the total surface. Non-linear areas of vegetation and bare soil are exceptional.

1.1 Urban fabric

1.1.2 Discontinuous urban fabric: Most of the land (Between 30 to 80 %) is covered by structures. Building, roads and artificially surfaced areas associated with vegetated areas and bare soil, which occupy discontinuous but significant surfaces. Areas mainly occupied by industrial activities of transformation and manufacturing, trade, financial activities and services, transport infrastructures for road traffic and rail networks, airport installations, river and sea port installations, including their associated lands and access infrastructures. Includes industrial livestock rearing facilities. 1.2.1 Industrial or commercial units Artificially surfaced areas (with concrete, asphalt, tarmacadam, or stabilised, e.g. beaten earth) without vegetation occupy most of the area, which also contains buildings and/or vegetation. Particularity of class 1.2.1: Agricultural farms: Areas of other than housing buildings, in-door spaces, stables, garages, workshops, lay-by and storing areas, often also bad land with ruderal vegetation, part of farms. The farms are often located in outskirts or close to rural settlements with agricultural function. Concentration of agricultural buildings in areas of various sizes was associated with collectivisation of agriculture. The quoted areas smaller than 25 ha are included in class 1.1.2. 1.2.3 Port areas: Infrastructure of port areas including quays, dockyards and marinas.

1.2 Industrial, commercial and transport units

1.2.4 Airports Airports installations: runways, buildings and associated land. 1.4.1 Green urban areas: Areas with vegetation within urban fabric, includes parks and cemeteries with vegetation, and mansions and their grounds

1.4 Artificial, non-agricultural vegetated areas 1.4.2 Sport and leisure facilities: Camping grounds, sports grounds, leisure parks, golf

courses, racecourses, etc. Includes formal parks not surrounded by urban areas. 2.4.2 Complex cultivation patterns: Juxtaposition of small parcels of diverse annual crops, pasture and/or permanent crops.

2.4 Heterogeneous agricultural areas

Particularity of class 2.4.2: Complex cultivation patterns with scattered houses Alternation of small plots (smaller than 25 ha) of arable land with annual or permanent crops with scattered garden huts or scattered houses. They are usually situated in proximity of rural or urban settlements and are used for growing agricultural crops, fruit, and vegetable for the particular households. Areas occupied by forests and woodlands with a vegetation pattern composed of native or exotic coniferous and/or deciduous trees and which can be used for the production of timber or other forest products. The forest trees are under normal climatic conditions higher than 5 m with a canopy closure of 30% at least. In case of young plantation, the minimum cut-off-point is 500 subjects by ha. 3.1.1 Broad-leaved forest: Vegetation formation composed principally of trees, including shrub and bush understoreys, where broad-leaved species predominate 3.1.2 Coniferous forest: Vegetation formation composed principaly of trees, including shrub and bush understoreys, where coniferous species predominate.

3.1 Forests

3.1.3 Mixed forest: Vegetation formation composed principally of trees, including shrub and bush understoreys, where neither broad-leaved nor coniferous species predominate.

EUFIRELAB

D-08-06.doc 81

Table 16b: Corine land cover nomenclature of the classes selected for the analyses (Combustible Land Covers)

CLC code CLC description

3.2.2 Moors and heathland: Vegetation with low and closed cover, dominated by bushes, shrubs and herbaceous plants (heather, briars, broom, gorse, laburnum, etc.) 3.2.3 Sclerophylous vegetation: Bushy sclerophyllous vegetation, includes maquis and garrigue. In case of shrub vegetation areas composed of sclerophyllous species such as Juniperus oxycedrus and heathland species such as Buxus spp. or Ostrya carpinifolia with no visible dominance (each species occupy about 50% of the area), priority will be given to sclerophyllous vegetation and the whole area will be assigned class 3.2.3.

3.2 Shrubs and/or herbaceous vegetation associations

3.2.4 Transitional woodland/shrub: Bushy or herbaceous vegetation with scattered trees. Can represent either woodland degradation or forest regeneration/recolonisation. Natural areas covered with little or no vegetation, including open thermophile formations of sandy or rocky grounds distributed on calcareous or siliceous soils frequently disturbed by erosion, steppic grasslands, perennial steppe-like grasslands, meso- and thermo-Mediterranean xerophile, mostly open, short-grass perennial grasslands, alpha steppes, vegetated or sparsely vegetated areas of stones on steep slopes, screes, cliffs, rock fares, limestone pavements with plant communities colonising their tracks, perpetual snow and ice, in land sand-dune, coastal sand-dunes and burnt areas.

3.3 Open spaces with little or no vegetation

3.3.4 Burnt areas: areas affected by recent fires, still mainly black.

Table 17: Sub-index 2 values according to the linear development of probable WUI areas (m). Natural breaks (Jenks) were selected to identify the six classes corresponding to the index values.

Sub-index 2 Linear development of probable WUI areas (m)

0 0

1 1-500

2 501-1000

3 1001-1700

4 1701-2700

5 2700-4700

Table 18: Sub-index 3 values according to the surface occupied in the cell by CLC2000 code 1.1.2. Natural breaks (Jenks) were selected to identify the classes corresponding to the index values.

Sub-index 3 CLC 112 surface (m2)

0 0

1 1-62.000

2 62.001 -200.000

3 200.001 - 380.000

4 380.001 - 650.000

5 > 650.000

Table 19: Final vulnerability index values and corresponding vulnerability classes.

Classes Vulnerability index

Low 1-5

Medium 6-10

High 11-14

EUFIRELAB

D-08-06.doc 82

9.2 VULNERABILITY RELATED TO ENVIRONMENTAL VALUE

9.2.1 Presentation

Ecological and landscape assessments are essential in the design of an integrated fire risk index which takes into account aspects related to the vulnerability of forested areas.

Such assessments are of great value in sustainable land management and help managers in decision making processes within a multipurpose scenario, ranging from natural environment conservation policies and biodiversity promotion strategies to urban and regional planning.

For instance, given the fact that the budget available for preventive silviculture is usually limited or insufficient, is it reasonable to apply these tasks to areas with the highest ecological and landscape value.

These areas are, therefore, the most vulnerable to this line of events.

In addition, when a forest fire is spreading, it would also seem reasonable for a fire manager to allocate fire suppression resources according to a certain scale of values.

This scale would bear in mind the presence of population, property (buildings, etc.) and also outstanding natural environments, such as natural parks, reserves, special protection zones, preserved landscapes, etc.

In some previous studies, this problem has been addressed by using GIS tools.

However, in some cases (Comunidad de Madrid, 2000) the only parameter which is taken into account to measure priorities in forest fire fighting, is the ecological value of the area according to the quality and the degree of protection of the forested ecosystems.

There is no assessment of the landscape and no landscape ecology indices are applied. In other cases (GULINCK et al. 2001), the method used to asses a given landscape is based on land use data, which does not include information from other supplementary data sources.

In this context, Landscape Ecology provides the appropriate conceptual framework that meets the demand for information which regional planners require.

It is well known that landscape ecology is a multidisciplinary science which mainly aims at solving the problems to do with land management and development on a local and regional scale (NAVEH and LIEBERMAN, 1994).

This new science facilitates territorial analysis by attempting to understand and compare different spatial patterns by means of patches with different shapes, quantities, classes, etc. (HONG et al. 2000; RIITTERS et al. 1995).

Landscape is under the influence of ecological and human processes, both on a regional and local scale, which imprint changes in its structure and composition.

The study of landscape structure/composition by means of spatial statistics helps us characterize the territory and understand the spatial-temporal relationship among the different elements that compose a landscape.

Landscape ecology has developed many indices that measure spatial textures and shapes, as well as the spatial structure of the landscape (MCGARIGAL et al. 2002; RIITTERS et al. 1995; MCGARIGAL y MARKS, 1995): patch density, size, compactness, fractal dimension, dispersion, diversity, etc.

Metric assessment and analysis of landscapes, using a GIS (MCGARIGAL et al. 2002; BERRY, et al. 1999; IGIS, 1997; MCGARIGAL y MARKS, 1995), allows to characterise the structures and changes in forested land occupation and land use within a territory, and also helps us find the environmental involvement of its activity (MALDENOFF y BAKER, 2000).

The studies carried out by Martinez et al., 2006 propose a simple method to map the landscape and ecological value of an area, by integrating several indicators in an exclusively ecological approach, without considering visual or aesthetic aspects.

The results show that the methodology can be usefully applied on a regional/global scale.

In addition, it is a valuable source of information for fire fighting authorities in decision making processes related to preventive forestry and to the distribution of extinction resources and manpower when a forest fire is spreading and threatening the areas which are most vulnerable from an ecological point of view.

We consider it would interesting to include the landscape assessment methodology proposed by these authors in the section concerning the vulnerability of an integrated risk index on a European level.

The methodology starts by integrating different indicators, under an exclusively ecological approach, and seeks to work out the intrinsic ecological value of a territory.

Figure 78 shows the methodological flow chart proposed by Martínez et al. (2006):

The main data source used to obtain the landscape value is the land cover map from the CORINE-Land Cover project from the year 2000 (CLC2000) available in digital format for all of Europe.

Two criteria groups are considered.

i) Vegetation and land use

The first group of indicators is based on the intrinsic characteristics of vegetation, such as the degree of proximity to climax vegetation.

In addition, it analyses the importance of vegetation types in its regional context (rarity) and in its global context (representativity).

ii) Landscape Ecology Indicators

As well as the variables from the previous group, other landscape ecology indicators are used, especially indices related to diversity, connectance and juxtaposition or interspersion, which take into account the spatial distribution of the patches within a territory.

EUFIRELAB

D-08-06.doc 83

MARTINEZ et al. (2006) propose a selection of indices (Simpson's Diversity Index, Interspersion Juxtaposition Index and the Connectance Index) with which to get to know the diversity, homogeneity or fragmentation of a given landscape.

Simpson’s Diversity Index (SIDI) assesses the number of different land cover types and the distribution of the area proportional the types of land covers.

Values near 0 indicate that there is only one patch or land cover (non diversity) whereas values close to 1 (high diversity) indicate that the different types of land covers as well as the proportional distribution of their areas is more balanced.

The Interspersion Juxtaposition Index (IJI) allows us to understand the spatial configuration of the patches as well as its contiguity and its degree of interspersion.

Low values (0) represent landscapes where patches are distributed randomly, whereas high values (100) correspond to landscapes which are distributed in an equal contiguity.

This index shows whether patches are grouped or distributed homogeneously in space.

The Connectance Index (CI) is defined by the number of functional unions among all the patches within the same type.

It show where each pair of patches are connected according to a 500 meter analysis window.

Values range from 0 (consisting in just one match or when none of the matches are connected) to 100, when all the patches in the analysis window are connected.

The Landscape Integrated Value (LIV) is the simple average of the combination of the six components mentioned.

An example of this product can be seen in figure 79.

Regarding the ecological value, the authors suggest considering whether the territory analysed is inside a preserved area of some kind: protected natural environments, special protection areas for birds, sites and habitats of community interest, preserved woodlands and public woodlands, etc.

This information can be easily accessed on a national/European scale.

In view of the results obtained in this work and although the workload capacity of GIS could include other elements for improving landscape and ecological assessment, we believe that the methodology proposed by the authors is straightforward and easy to understand.

In addition, it has basic data requirements, available on a European scale, that guarantee the applicability of the method and its integration, together with other factors, into a synthetic fire risk index.

9.2.2 Figures

Figure 78: Landscape assessment methodological flow chart

EUFIRELAB

D-08-06.doc 84

Figure 79: Landscape integrated value map for the Madrid region (MARTÍNEZ et al., 2006)

EUFIRELAB

D-08-06.doc 85

9.3 POTENTIAL SOIL EROSION

9.3.1 Introduction

Soil erosion is a widespread threat to European soils.

The main consequences of erosion are not only on-site soil degradation, declining fertility, desertification and reduced infiltration and water storage capacities, but there are also off-site impacts that include eutrophication of rivers and lakes, destruction of wildlife habitats, siltation of dams, reservoirs, rivers, as well as infrastructure and property damage by muddy floods (RUBIO et al., 2006).

Some authors have highlighted its impact on global food security (CROSSON, 1997; LOMBORG, 2001).

In addition, soil erosion results in emission of soil organic carbon to the atmosphere in the form of CO2 and CH4, thereby enhancing global warming (LAL, 2004).

Global warming in turn is expected to increase erosion rates (NEARING et al., 2004).

European Research Programmes initiated by the European Commission demonstrate that although the Mediterranean region is historically the most severely affected by erosion, there is growing evidence of significant erosion occurring in other parts of Europe (e.g. Austria, Czech Republic and the loess belt of Northern France and Belgium).

Soil erosion can therefore be considered, with different levels of severity, an EU-wide problem (MONTANARELLA et al., 2003).

Soil erosion occurs under diverse conditions and is driven by interactions of many factors such as land use, climate, soil conditions, and topography that are difficult to quantify.

A proper assessment of erosion problems is greatly dependent on their spatial, economic, environmental, and cultural context (WARREN, 2002).

As a result, a comprehensive understanding of soil erosion is still very difficult.

Forest fires are a further factor that can lead to excessively high rates of erosion and contribute to environmental degradation.

Every year there are more than 50,000 forest fires in Europe, affecting over 500,000 ha of forest and other woodlands, the majority of which occurs in the Mediterranean Region (European Commission, 2002).

The continuous impacts of fires, with progressively shorter recovery periods, together with torrential rains characteristic of the Mediterranean climate, favour the intensification of erosive processes (GIOVANNINI et al., 1990).

Wildfires and prescribed fires affect the vegetation, soils, wildlife, and water resources of watersheds.

They impose a wide range of effects which depend on the mosaic of fire severity and post-fire hydrologic events (NEARY, 2004).

Research indicates that fires are likely to increase runoff rates and sediment yields relative to undisturbed forested land (AHLGREN and AHLGREN, 1960; LAIRD and HARVEY, 1986; May, 1990; SCOTT and VAN WYK, 1990; SOTO et al., 1991; SOLER and SALA, 1992; SWANSON, 1981).

A great number of pan-European soil erosion risk assessment efforts have been made with a variety of approaches such as GLASOD approach (Global Assessment of the Current Status of Human-Induced Soil Degradation), INRA (Institut National de la Recherche Agronomique, 1988) approach, HOT-SPOTS, IMAGE/RIVM, CORINE, USLE/ESB, PESERA.

Widely used models include RUSLE (Renard et al., 1997) and MUSLE (SMITH et al., 1984).

Another model, Soil Erosion Model for Mediterranean regions (SEMMED, DE JONG, 1994a) was developed with the objective of using satellite data, similarly to the Thornes model (THORNES, 1985), Agricultural Nonpoint Source Pollution model (AGNPS: YOUNG et al., 1989) and the Areal Nonpoint Source Watershed Environment Response Simulation model (ANSWERS: BEASLEY et al., 1980).

Furthermore, important European policies and directives, such as the Water Framework Directive (EC, 2000), the European Commission Soil Thematic Strategy (COM, 2002), and instruments of the Common Agricultural Policy, such as agro-environmental measures (EC, 1999), address the issues of soil erosion.

Although there are many approaches in relation to soil erosion risk assessment, only a few can be found in the literature that take into account the effects of wildfires on vegetation and soil conditions and are suitable for burned woodland areas.

Most models and strategies address soil erosion problems in agricultural areas only.

9.3.2 Objectives

In accordance with the aim of this deliverable, a vulnerability index of soil erosion will be created for Mediterranean countries that suffer high intensity and frequency of forest fires.

The questions and objectives that were set regarding potential soil erosion in the previous deliverable (D-08-03) are the following: - To what extent can planners prepare for the

consequences of a forest fire before it occurs in order to respond more quickly to the potential runoff and erosion risks

- To elaborate a pre-fire strategy that would shorten the post-fire reaction time

- To evaluate the use of the model quoted in D-08-03 (ANNEX I) at different spatial scales and to answer how model scale affects the objectives and the operational use of the model.

EUFIRELAB

D-08-06.doc 86

From the above questions, two approaches can be put forward.

In the first, a “what if” logic underlies the model in order to define potential erosion risk areas.

The underlying question is therefore “what would be the post-fire soil erosion risk if a forest fire were to take place? (sub-chapter 9.3.3)”.

In this case, scenarios can be tested at the European scale to define potential soil erosion risk.

The focus here is on gaining a spatial perspective of the distribution of high-risk zones.

In a second approach, an operational model is elaborated to be used in a Mediterranean context in the case of an actual fire.

Here, the question is “how can Mediterranean countries best respond to a fire in order to deal efficiently with the erosive consequences? (sub-chapter 9.3.4)”.

9.3.3 A European scale model to predict potential erosion risk

The aim of this approach is to perform a pre-fire prediction of post-fire soil erosion risk.

Post fire soil erosion risk is strongly related to the effects of fire on vegetation and soil conditions.

In other words, before a fire actually occurs, we must firstly predict the kind and intensity of fire effects on soil properties and vegetation; secondly, we must be able to estimate the consequences of these changes on soil erosion rates.

One approach to addressing this problem is to modify the method quoted in D-08-03 deliverable (ANNEX I) in order to be able to acquire information about the post-fire erosion risk that may arise before a Wildfire takes place.

This ‘a priori’ knowledge would give European level planners a vision of the spatial distribution of post-fire erosion risk and help in the preparation of strategies at the European program scale to deal with this potential danger.

9.3.3.1 Discussion and suggestions about the method described in ANNEX I

The method quoted in D-08-03 deliverable takes into account four factors to produce a map of soil erosion risk at the catchments scale (FOX et al., 2006).

These factors are: - Slope - Pre-fire vegetation - Fire severity - Soil erodibility

These factors were estimated after a wildfire in the Massif des Maures area near St Tropez in southern France which means that this method poses temporal difficulties in the typical Mediterranean climatic context where there is a limited time between a summer or autumn fire and the following rainy winter period.

As a result, this method has to be modified in order to meet the following needs: - to be applied in pre-fire time and thus help us to

acquire an ‘a priori’ knowledge of the most soil erosion prone areas

- to provide European scale information of the potential erosion risk.

The way the four erosion factors were estimated and how they can now be addressed in order to predict potential soil erosion at the European scale in pre-fire time is discussed below.

9.3.3.2 Slope

This factor can be considered static since it does not change over time.

A 25-m DEM was used for slope estimation, and a 50-m resolution was considered too coarse to be operational.

At the European scale, 30 to 50-m DEMs are available and are considered sufficient of the objectives of this approach.

Slope needs to be estimated at the European scale only once since it does not change significantly over time.

FOX et al. (2006)formed five categories, according to slope magnitude and coefficients were ascribed to each category as follows:

9.3.3.3 Pre-fire Vegetation density

This factor was estimated from the standing charred trunks during a field survey and three categories were included.

Coefficients were ascribed to each factor level as follows table 20:

It is obvious that at the European scale, vegetation data cannot be collected manually.

However, CORINE land cover or remote sensing data could be used to estimate pre-fire vegetation.

Vegetation cover provides protection of the soil against erosion processes.

This happens not only before fire (INBAR et al., 1997), but also after fire because in low intensity and medium intensity fires the surface organic matter is increased providing protection from soil erosion (fallen needles, twigs, etc).

In addition to the mulch effect provided by fallen pine needles (SHAKESBY et al., 1993; SHAKESBY et al., 1994), standing dead vegetation may favour infiltration (near stems and in burned roots) and slow runoff velocity, so the net expected effect is lower erosion where vegetation stands are denser.

ANDREU et al. 1996 demonstrated that organic matter content strongly decreases immediately after medium to intense fires, with a subsequent increase with time.

However, in zones that suffered a low to medium fire intensity its value increases.

This could be due to the accumulation of plant residues, not completely burnt, according to the studies of CHRISTENSEN 1987.

EUFIRELAB

D-08-06.doc 87

It was found from the literature that for land cover estimation, the following methods were applied: - in the IMAGE/RIVM approach, Olson’s land cover

database and statistical information from FAO was used

- in the CORINE approach, the CORINE landcover database was used

- in the USLE/ESB approach, an exponential scaling function was linked to the Normalized Difference Vegetation Index (NDVI) extracted from NOAA images

- in the PESERA approach, a plant growth model was used initially, but in monitoring mode it was replaced by NDVI from Earth Observation Sensors NOAA-AVHRR or SPOT-VGT (also applied by Zhang et al., 2004)

- an NDVI (TUCKER et al., 1985) was also extracted from Landsat TM data (Bayaramin et al., 2006) in the semi-arid area of Beypazari in Ankara in order to create a layer of vegetation cover which was used in a study of soil erosion risk assessment. NDVI was grouped as “fully protected” or “not fully protected” based on ground truth information with the global positioning system. The map created can be seen further down (Figure 80). The pixel size from NOAA-AVHRR is of 1 km spatial resolution so in order to improve the spatial resolution, LANDSAT remote sensing data could be used in local scale to calculate Vegetation Indices (VIEDMA et al., 1997: Santos et al., 2000: DELGADO et al., 2001) in which pixel size is 30 m. This can easily be used with a 30-m DEM.

Resolution of 30 by 30-m pixel size can be considered appropriate because of the great fluctuation in vegetation density that can be found inside a 1 by 1 km pixel size of NOAA-AVHRR.

Although LANDSAT remote sensing data could be used for NDVI calculation, it is not convenient for European scale projects because of its swath width (185 km).

The Disaster Monitoring Constellation (DMC) can be used for NDVI calculation.

This system records reflected electromagnetic radiation over the same spectral range as Landsat-ETM 2, 3 and 4 bands, it offers 32 m spatial resolution, daily coverage and a very wide swath of 600-km. NDVI data can be used in combination with thresholds in order to classify the pixels into the three aforementioned categories of land cover density.

In addition, a fourth category, high vegetation density, can be added to the table above.

This class was not encountered in the Massif des Maures catchment described in deliverable D-08-03, but would be found at the European scale.

Finally, it is worth mentioning that the DMC is the first earth observation constellation of five low cost satellites providing daily images for applications including global disaster monitoring.

9.3.3.4 Fire severity

It must be defined as potential fire severity.

Fire severity was defined as either variable or intense and coefficients were ascribed to each level as following: Table 22

“Variable severity” were sites where either the fire had passed quickly and charred the trunks and underbrush, but left the crown more or less intact, or the intense combustion of the standing vegetation was confined to a small area (dozens of m2).

The “intense severity” category consisted of sites where the fire had consumed the underbrush and crown entirely over an extensive area.

According to this classification, this factor could also be defined as fire type because the former category corresponds more or less to the description of a surface fire, whereas the latter to the description of a crown fire. In order to estimate fire severity, the main factors contributing to the outbreak of a crown fire should be taken into account.

The primary factors influencing crown fuel ignition are: - Height of the under-story vegetation and - The vertical distance between the ground/surface

fuel strata and the lower boundary of the crown fuel layer (Miguel et al., 2006).

These two factors are the most significant for predicting danger of crown fire ignition.

Although, they cannot be estimated easily at the European scale from remote sensing data, because methods that can be applied for their estimation, are usually, if not always, by means of LIDAR data (ANDERSEN et al., 2005: DRAKEA et al., 2002: MORSDORF et al., 2004: RIANO et al., 2003a).

This system is airborne and an inventory effort at such a scale would be practically and financially unfeasible.

Instead of this method, it is proposed to initially reclassify vegetation types into the two aforementioned categories (variable or intense fire severity) based on the following: - vegetation types from the CORINE database in

relation with the potential existence of a dense understory below specific dominant tree species

- the flammability of species (Dimitrakopoulos and Papaioannou, 2001)

It is widely known that some dominant tree species are associated with specific densities and structures of understory vegetation, as is the case for Pinus halepensis where we find very dense underbrush in contrast to Fagus species under which there is limited understory vegetation.

Flammability of species on the other hand can help us to predict the intensity and severity of a forest fire before it occurs.

EUFIRELAB

D-08-06.doc 88

The periodicity of estimation for this factor could follow that of the Propagation Danger Index described elsewhere in this deliverable since they are highly correlated.

Other factors that can lead to high severity wildfires are the high density and favourable horizontal structure of vegetation cover which may result in extensive spreading, low live and dead fuel moisture contents, intense winds, high temperatures and low relative humidity.

The vegetation cover density could be assessed as stated earlier with NDVI calculation whereas information about the live and dead fuel moisture content from the corresponding indices developed in other sections of this deliverable.

Wind speed, temperature, and relative humidity can be estimated either from historical meteorological data or from meteorological stations across Europe.

9.3.3.5 Soil erodibility

Three classes of soil conditions were defined and coefficients were ascribed to each class as following: Table 23

Shallow stony soils are assumed to generate more runoff but surface material is less mobile due to the high stoniness, and in many cases, the bare bedrock.

In contrast, deep soils with fewer stones may have higher infiltration rates, but they are more erodible, particularly with respect to the weak aggregate stability and possible presence of hydrophobicity.

Therefore, it was decided to increase the value of the coefficient with decreasing stone content.

In order to estimate this factor, without altering the logic upon which it was built and its functionality, it is proposed to use the method applied in the CORINE approach: erodibility is estimated from soil texture, depth and stoniness, extracted from the soil map of the European Communities (CEC, 1985).

The CORINE model, which has been applied to many countries in the European Community, is constituted by combining 4 parameters: soil erodibility, erosivity, topography and vegetation cover (DENGIZ and AKGUL, 2005).

The following diagram (Figure 81) shows the three parameters used to estimate erodibility and how they can be integrated into four classes.

Figure 82 (taken from KIRKBY, 2001) shows the overall flow chart of CORINE model for risk assessment of soil erosion

In the IMAGE/RIVM approach, terrain erodibility was estimated based on soil type and landform.

Landform was classified into types by using the difference between minimum and maximum altitudes for each grid cell whereas soil type: - was derived from the FAO Soil Map of the World

and - is composed of soil depth, soil texture, and bulk

density (GOBIN et al., 2003).

The disadvantage of this approach is that the information derived is at a 50-km spatial resolution, which renders it difficult to interpret at sub-national scales (GOBIN et al., 2003).

Soil type can be considered a very important factor affecting erodibility since some soil types present higher susceptibility under the same conditions.

For example, PAPAMIXOS 1996 states that terrarosa soils present higher soil erosion risk in comparison to other soil types, especially for high slope values.

Studies of GIOVANNINI et al. 1990 indicate that, depending on the fire intensity; CaCO3 could suffer a break-up and change to CaO or Ca+2 (ANDREU et al., 1996).

This new constitution is far more erodible and can be washed away by water more easily.

GOFAS (2001) refers also to.

It has been mentioned in the D-08-03 deliverable that high surface temperatures volatise organic materials and create gases that move downward in response to a temperature gradient and then condense on soil particles causing them to become water repellent (LETEY, 2001).

This causes an increase in surface runoff and therefore soil erosion rate.

Many researchers focus on the changes in soil properties as the essential element affecting erosion rates (IMESON et al., 1992; KUTIEL and INBAR, 1993; SEVINK et al., 1989).

GIOVANNINI and LUCCHESI 1991 stress the importance of the soil conditions as the main factor in the erosion processes.

In soils with a high percentage of organic matter, intense fire may lead to complete destruction of the organic layer and the exposure of a hydrophobic mineral soil layer, which acts as a repellent mantle (DEBANO, 1981).

This increases runoff and erosion rates (SEVINK et al., 1989).

IMESON et al. 1992 state that for fires of medium and high intensity, the degradation of the vegetation cover and soil organic matter produce the surface accumulation of hydrophobic substances that reduce infiltration and increase runoff (ANDREU et al., 1996).

This situation causes increases in runoff production and soil removal, which favour the loss of nutrients.

The severity of water repellency depends on the combined interactions of soil properties and the soil-heating regime developing during a fire.

The longevity of fire-induced water repellency depends on some of the same factors that affect its formation.

Water repellency produced by low to moderate severity fires is usually of shorter duration than that produced by high severity fires (DEBANO, 2006).

The results of preliminary field observations suggest that water repellency might well be an important factor responsible for the accelerated erosion experienced during the first few years following wildfires (KRAMMES and DEBANO, 1965).

EUFIRELAB

D-08-06.doc 89

An initial laboratory study showed that water repellency could be intensified by heating a soil - organic matter mixture in a muffle furnace at different temperatures for different lengths of time (DEBANO and KRAMMES, 1966).

It was hypothesized that a more efficient coating of mineral soil particles occurred at lower temperatures and for shorter periods of heating than in the case of longer periods of heating at higher temperatures, which destroyed the organic substances responsible for the water repellency.

High intensity fires, of 1,200 °C or more, do not always result in high severity impacts on the soil if their duration is short, but low intensity fires of just 300 °C that smoulder for a long time in roots or organic matter can produce large changes in the nearby soil (NEARY, 2004).

The variability of erosion is associated with fire severity. Fire severity has been reported (INBAR et al., 1998; ROBICHAUD, 2000) as being one of the main reasons for causing erosion variability.

Fire effects on soil depend mainly on fuel type, fire severity/intensity, topography and soil conditions (PAPAMIXOS, 1996).

Nevertheless, the way soil erodibility is addressed in this method can be considered as simplified since some of the factors affecting susceptibility (soil type and wildfire effects on soil properties) are not taken fully into account.

However, extracting this type of data before an actual event, and even after an event without direct measurements, is unfeasible at this time.

9.3.3.6 Conclusions

In the Mediterranean, heavy rains follow the fire season and this means that intervention measures have to be undertaken upon urgently.

Up until today there is a number of factors, which are related to soil erosion risk assessment that are difficult to estimate.

The proposed methodology is based on the use of factors that are easy to estimate either directly or indirectly.

It constitutes an approach for “fast and coarse” risk assessment of potential soil erosion after a fire event.

It should be mentioned that Soil Erosion is closely related to the effects of fire on vegetation and soil.

The proposed methodology still needs to be evaluated by conduction of extensive field surveys after a fire event.

9.3.4 An operational model for the Mediterranean context

The post-fire effects of a wildfire are usually temporally and operationally problematic because the first winter following a summer forest fire is the most vulnerable period, (SOTO and DÍAZ-FIERROS, 1998; Vacca et al., 2000), and runoff and erosion management strategies have to be implemented in the first weeks after a fire.

The techniques used and their spatial location must often be decided upon urgently and in an administrative context where the sources of funding, funding channels and agency responsibilities are all evolving over a period of days to months (FOX et al., 2006).

In deliverable D-08-03, an operational model was presented to respond to post-fire runoff and erosion risks.

A full description of the model can be found in Appendix 1.

Initial testing of the model proved satisfactory, but an additional procedure has been introduced to improve it further still.

The model itself will not be revised here since it has been the subject of discussion in the preceding section and can be consulted in full in the appendix.

The following section deals only with the improvement to the model and it resides in its capacity to estimate the evolution in erosion rate over the first years following a fire.

The change in erosion rate depends principally on vegetation re-growth, and this in turn is closely related to soil properties.

In the Mediterranean environment, both soil properties and vegetation characteristics are strongly influenced by topography.

The objective was therefore to predict temporal changes in erosion rate over a 6 year period based on the impact of topography on soil properties and vegetation growth.

The proposed addition to the model was developed from a detailed study of the relationships between topography, landuse, soil properties, and post-fire vegetation re-growth in the same catchment used to test the initial model.

Although the relationships from this single catchment cannot be extrapolated to the entire Mediterranean region, they are typical of many Mediterranean environments.

EUFIRELAB

D-08-06.doc 90

9.3.4.1 Landuse versus topography

In Mediterranean regions, there is a distinct relationship between landuse and topography.

The climate is conducive to the production of olives, vines and other crops.

In Mediterranean coastal areas of high population density, suitable land, possessing both relatively low slopes and fertile soils, are shared between agricultural production and human occupation.

This is clearly illustrated in Figure 83, where the valley bottoms are occupied almost entirely by vineyards or urban development and the steeper upland areas are forested with some individual housing.

The spatial distribution in landuse has implications for post-fire soil erosion risks.

In Mediterranean France and throughout much of the Mediterranean region, forest fires occur systematically on the steepest slopes of the catchment.

The post-fire context is therefore naturally highly erosive where slopes tend to be both steeper and longer (Figure 84): slope length in the field is determined essentially by field length, whereas in forested areas it is not uncommon to find slopes that extend from the crest down into valley bottoms where slope inclination levels off.

The impacts of slope inclination and length on erosion rates are well known and are not elaborated upon here.

In addition to the impact of slope inclination and length on erosion rate, the steeper slopes also influence sediment deposition.

Sediments detached from slopes have a low probability of being deposited before reaching a tributary, and, once in the tributary, there is little likelihood of deposition before they reach the main channel.

As can be seen in Figure 85, the area that could serve as a potential zone of deposition is much more restricted in the forested zone.

This is quite different from an agricultural context where fields along river channels can serve as sinks for sediments eroded upslope and where channel deposition can be important.

Therefore, the relationship between landuse and topography affects not only erosion rate, but it also has an impact on sediment redistribution.

9.3.4.2 Topography versus soil characteristics

The relationships between topography and soil characteristics were determined by an analysis of the impacts of topographic units on soil properties.

Slopes were divided into three categories: top-slope convexities, North-facing slopes and South-facing slopes. In many Mediterranean environments, water is a limiting factor and soils tend to be deeper and better developed on North facing slopes where evapo-transpiration rates are lower.

In this study, two soil properties were taken into consideration: soil depth, measured using a hand-held soil corer, and grain size distribution.

Textural analyses were carried out on surface (top 5 cm) samples for each topographic section and these were compared to sediments accumulated in sediment traps put in place after the 2003 fire described in the previous deliverable.

Thirty sites were selected, ten on each slope section.

Figure 86 shows the distribution of soil depths according to slope section.

North facing slopes clearly have deeper soils (α=0.05, r2=0.74 in an analysis of variance).

Soil depths on these soils since measurements were stopped at a depth of 1 m due to the difficulties in manipulating the hand held corer at greater depths.

Differences in soil depth arise essentially from long-term soil forming processes.

Better moister conditions on North facing slopes increase the rate and duration of pedogenic processes and favour vegetation growth (described below).

Greater rooting depths and vegetation densities, in turn, contribute to more active soil formation.

In addition to differences in soil depth, we can also note changes in surface grain size distributions.

The stacked histogram of Figure 87a shows that the fine fraction (< 0.2 cm) content on the South facing and convex slopes is less than half the content on the North facing slopes.

The sediments trapped in sediment traps (Figure 87b) positioned in a stream channel draining a burned slope (see FOX et al., in press) can explain this tendency.

Log Debris Dams (LDD) are located in the channel and tend to trap coarser sediments.

Sedimentation basin (Basin) sediments represent the fraction of sediments moved through the channel after initial settling behind the LDDs.

Figure 87b clearly shows selective erosion where the finer fraction is removed preferentially from burned slopes.

The results suggest that post-fire erosion processes occur longer and/or more frequently on South facing and convex slopes.

Finally, it should be noted that the gravel size fraction (0.2-0.5 cm) represents a greater fraction of the surface texture on convexities than on South facing slopes.

In this case, runoff on the low inclination convexities has sufficient velocity remove the finest fraction, but not the coarse (> 0.2 cm) sediments, as is the case on the South facing slopes.

EUFIRELAB

D-08-06.doc 91

9.3.4.3 Topography versus vegetation

For each of the 3 plots identified, vegetation characteristics were measured in two ways.

Vegetation height was measured on a random sample of 20 plants located in a 2 m by 2 m perimeter.

This was carried out by hand using a tape measure. Vegetation density was quantified by photographing

a 50 cm by 50 cm grid (Figure 88) and then by noting the percentage of points occupied by vegetation in a random sample of 50 points.

For each x,y coordinate in the 50-points sample, the surface was noted as either vegetated (1) or non-vegetated (0).

This value was then converted to a percentage. Three photos were taken at each of the 30 plots for

a total sample size of 90-grid pictures.

As expected, both vegetation height and density were greater on North facing slopes than South facing slopes and top-slope convexities (Figure 89).

Furthermore, soil depth, vegetation height and surface cover were all positively correlated.

Slope was significant in analyses of variance (α=0.05) for both vegetation height (r2=0.74) and surface cover (r2=0.63).

The greater vegetation growth on North-facing slopes supports the hypothesis evoked above that selective erosion is more intense on South facing slopes than North facing slopes since the latter take more time to recover after a fire.

In addition, South-facing slopes have always had more human occupation, as is shown by old terraces and newer housing, so the likelihood of fire ignition has always been greater on South facing slopes.

This, combined with drier conditions, suggests that South-facing slopes may burn more often than North facing slopes.

9.3.4.4 Implications of topographic effects for erosion modelling

The model described in Annex gives an instantaneous spatial distribution of post-fire soil erosion risk.

However, post-fire erosion, though at its maximum during the first year, can continue over a period of several years.

Faster vegetation re-growth on North facing slopes suggests that erosion will be decrease more rapidly here than on South-facing slopes or top-slope convexities.

This is supported by both the vegetation data gathered in-situ and the soil and textural results.

To complete the model, it is therefore necessary to consider this temporal evolution.

Figure 90 shows the initial calculation of the index of soil erosion risk.

The initial model re-classed these values into categories according to erosion risk.

It should be noted that these values represent a post-fire structural risk.

Actual erosion rates depend on random rainfall effects: for the same erosion index, erosion rates can be significantly different according to the total rainfall and number of high rainfall intensity events.

Hence, an erosion index of 90 (maximum value in Figure 90) can correspond to an actual erosion rate of from a few T ha-1 to dozens of T ha-1 according to post-fire rainfall patterns.

In addition to the impact of rainfall described above, several processes contribute to determine the evolution in post-fire erosion.

At least three can be described here: - The initial rates of erosion depend on the magnitude

of the first post-fire rainfall events as described above.

- The initial decrease in erosion rates can be attributed to the loss of the most easily mobilised sediments. As the finer fraction is eroded, the increasing stone cover at the surface progressively decreases sediment detachability and erosion rates.

- Erosion rates decrease further due to the progressive establishment of a vegetation cover on the slopes. This occurs faster on North facing slopes than South-facing slopes.

The trends cited above are shown in figure 91. The decrease in the erosion index trend is different

for North-facing and South-facing slopes since vegetation dynamics proceed at different rates.

The change in erosion index over time was estimated using non-linear decaying exponential curves for each of the slope types (Table 24).

The two extremes, North- and South-facing slopes, were defined so that the index reached approximately zero within three years for the North-facing slope and within six years for the South-facing slope.

The 3-year limit for the North-facing slope was determined in the field by observations of the filling of a sedimentation basin.

The 6-year limit for the South-facing slope is an estimation based on the current vegetation cover.

Convex slopes were treated with the same equation as South-facing slopes and West- and East-facing slopes were given decay values intermediate to the two extremes.

Slope orientation was defined for each cell in Figure 90.

Top-slope convexities were defined as slopes above 150-m with an inclination of less than 10%.

This proved to be a simple but effective method of identifying these topographic features.

Once the orientation was simplified into the five categories in table 24, the appropriate regression equation was applied to each of the cells, year by year for 6 years, in order to predict its evolution over time.

For each year, the erosion index value was converted to an erosion risk class, as in the initial method (see Annex).

The result is a dynamic erosion risk spatial distribution showing the potential erosion risk for each year during the 6-year period (Figures 92a to 92f).

EUFIRELAB

D-08-06.doc 92

The map for the sixth and final year is not shown since the entire surface is in the low class category.

In the initial year after the fire, the North facing slopes initially in the very high-risk class have dropped to the high and intermediate classes.

By the third year, all of the North-facing slopes are in the low risk class.

Many of the South-facing slopes persist in the very high-risk class after the first year, despite the exponential decrease in erosion risk, and many of these slopes remain in the intermediate class even three years after the fire.

Showing the temporal evolution in erosion risk is an added factor in determining where erosion control measures should be put into place after a fire.

9.3.4.5 Conclusions

The Mediterranean environment is subject to regular forest fires and it is possible that the number and intensity of fires will increase in coming years.

Reacting quickly to post-fire erosion risk is therefore an issue that must be dealt with by European countries in particular.

The above deliverable demonstrated two approaches to dealing with post-fire erosion risk.

In the first, a method for determining pre-fire potential risk at the European scale is described.

In this context, the issue addressed is how to acquire apriori knowledge about post-fire erosion risk before an actual fire occurs.

The second approach refines a post-fire erosion mapping method that was presented in deliverable D-08-03.

The major improvement concerns the integration of post-fire erosion risk evolution over a period of 6 years after the fire.

In this approach, it was demonstrated that topography plays a major role in determining post-fire vegetation recovery and soil depth, and these, in turn, have a major influence on post-fire erosion.

Field data support the approach and it has considerable potential for testing in a number of Mediterranean post-fire contexts.

9.3.5 Figures

Figure 80: Map of NDVI produced by Bayaramin et al. (2006) for their study area.

EUFIRELAB

D-08-06.doc 93

Soil texture (ST) 0 for Bare rock 1 for C, SC, SiC 2 for SCL, CL, SiCL, LS, S 3 for L, SiL, Si, SL

Soil Depth (SD) 1 for > 75 cm 2 for 25-75 cm 3 for 25 cm

Stoniness (SS) 1 for >10 % 2 for <10 %

Soil erodibility(K) 0 for ST*SD*SS=0 1 for 0 < ST*SD*SS < 3 2 for 3< ST*SD*SS < 6 3 for ST*SD*SS > 6

Figure 81: Methodology for CORINE soil erodibility assessment (CORINE, 1992).

Soil Texture, ST

0 for Bare rock 1 for C, SaC, SiC 2 for SsCL,CL,SiCL,Lsa 3 for SaL, L, SiL, Si

Soil Depth, SD 1 for >75 cm 2 for 25 – 75 cm 3 for <25 cm

Soil Stoniness, SS 1 for >10% 2 for <10%

Erodibility, K 0 for ST.SD.SS =0 1 for 0< ST.SD.SS <32 for 3< ST.SD.SS <63 for ST.SD.SS >6

Fournier Index, F 1 for Σpi

2/Σp <60 2 for 60<Σpi

2/Σp<90 3 for 91<Σpi

2/Σp<120 4 for Σpi

2/Σp>120

Bagnouls-Gaussen Aridity Index, B 1 for Σ(2Ti-pi) =0 2 for 0<Σ(2Ti-pi)<50 3 for 50<Σ(2Ti-pi)<130 4 for Σ(2Ti-pi)>130

Erosivity, R 1 for F.B<4 2 for 4<F.B<8 3 for F.B>8

Potential Soil Erosion Risk, EP 0 for K.R.S = 0 1 for 0 < K.R.S < 5 2 for 5 < K.R.S < 11 3 for K.R.S > 11

Slope angle, S 1 for <5% 2 for 5-15% 3 for 15-30% 4 for >30%

Actual Soil Erosion Risk, EA 0 for EP.V = 0 1 for EP.V = 1-2 2 for EP.V = 3-4 3 for EP.V >=5Land Cover, V

1 for fully protected 2 for not fully protected

Figure 82: Methodology for CORINE Soil Erosion assessment

EUFIRELAB

D-08-06.doc 94

Topographyand

Landuse

Alluvial plainHilly metamorphicuplands

ForestVineya rdsOtherUrban

Figure 83: Steeper uplands in the coastal area of the Massif des Maures (France) are occupied by forest, the

lowland plains are occupied by vineyards and urban development.

0 20 40 60 80 100Slope Inclination (%)

0

7

14

21

28

35

Area per Landuse Type (%)

Non-forestForest

Mean (%) Median (%)

Forest: 23.3 25.5

Non-forest: 10.1 6.6

a) Slope inclination b) Slope length

Forest Vineyard0

100

200

300

400

Slope Length (m)

Mean (m) Median (m)

Forest: 357.1 400.0

Non-forest: 162.7 162.0

Figure 84: Highly erosive topographic conditions with few possibilities for sediment deposition in the fire zone

EUFIRELAB

D-08-06.doc 95

Forest Non-forest0

10

20

30

40

50A

rea

( %)

North South Convex0

20

40

60

80

100

Soi

l Dep

th (c

m)

Left Figure 85: Percentage area in the forest and non-forest zones with a slope inclination less than 4%. Right Figure 86: The North-facing slopes have a greater depth than the South-facing and convex slopes.

a) Slope grainsizecharacteristics

b) Sediment trapgrainsize characteristics

North South Convex0

20

40

60

80

100

Grainsize Distribution (%)

>2 cm1-2 cm0.5-1 cm0.2-0.5 cm< 0.2 cm

BASIN LDD0

20

40

60

80

100

Initial Grainsize Distribution (%)

> 20 mm2-20 mm< 2.0 mm• Selective erosion of the finer fractions (< 2 mm) from slopes.

• Increasing stone cover with erosion – South and Convex slopes.

• Intermediate fraction (0.2-0.5 cm) detached from South facingslopes but not convexities

0.2-2 cm

Figure 87: North-facing slopes have finer textures than South facing and convex slopes (a) This results from selective erosion of the finer fraction (measured in sediment traps - b)).

EUFIRELAB

D-08-06.doc 96

Figure 88: Example of a grid photograph used to measure vegetation density.

North South Convex0

40

80

120

Vegetationheight (cm)

a) Undergrowth vegetation height b) Undergrowth vegetation cover

R2=0.74

North South Convex0

20

40

60

80

100

Vegetation cover (%)

R2=0.63

Correlations: Vegetation height/area: r=0.82; Soil depth / Vegetation height: r=0.87; Soil depth / Vegetation area: r=0.75

Figure 89: Vegetation height (a) and density (b) were both greater on North facing slopes than on South facing

slopes and top-slope convexities.

EUFIRELAB

D-08-06.doc 97

Theoretical range: 1-135; Actual range: 1-90

Figure 90: Raw values for soil erosion risk calculated from the initial soil erosion model.

0 1 2 3 4 5 6 7Time (years)

0

30

60

90

Initi

a l E

rosi

on In

d ex

0 1 2 3 4 5 6 70

30

60

90Peak erosion = f(rainfall events)

Initial decrease = f(sedimentavailability, increased stoninessthrough selective erosion)

Greatest decrease = f(vegetationgrowth)

North:

South:

Figure 91: Theoretical curves for soil erosion index trends over time: rates decrease and the greatest decrease arises due to differences in vegetation growth, which are significantly different on North and South facing slopes

EUFIRELAB

D-08-06.doc 98

Figure 92a and b: Initial soil erosion (left) and one year after the fire risk maps

Figure 92c and d: Erosion risk maps two (left) and three (right) years after the fire

Figure 92e: Erosion risk maps four (left) and five (right) years after the fire

EUFIRELAB

D-08-06.doc 99

9.3.6 Tables

Table 20: Coefficients assigned to each slope class

Slope (%) Coefficient

0-5 1

5-10 2

10-20 3

20-30 4

>30 5

Table 21: Coefficients assigned to each vegetation density

Pre-fire vegetation density Coefficient

Medium 1

Low 2

Bare 3

Table 22: Coefficients assigned to each fire severity level.

Fire severity Coefficient

Variable 1

Intense 3

Table 23: Coefficients assigned to each soil condition class

Soil erodibility Coefficient

Shallow soil with high stone content (depth < 20 cm) 1

Intermediate depth and stoniness 2

Deep soil with few stones (depth > 50 cm) 3

Table 24: Non-linear equations used to predict the evolution in the erosion index

Slope orientation Non-linear equation

North yreIVEI 2−×=

South yreIVEI 8.0−×=

Convex yreIVEI 8.0−×=

East yreIVEI 2.1−×=

West yreIVEI 6.1−×=

EUFIRELAB

D-08-06.doc 100

10 EURO-MEDITERRANEAN WILDLAND FIRE RISK INDEX

10.1 GENERAL PRESENTATION

The basic tools for the application of the proposed EM-WFRI for the regional wildfire management are available once the basic structure has been defined and the different components of this index have been identified.

Next, the methodology to integrate these components and finally the most suitable procedure to spread the information should be defined.

Concerning this, the multiple suggested possibilities in the D-08-05 must be remarked.

Given the diversity of available methods, and the variety of potential applications and users, whether of the total of a part of the identified risk components, a unique integration methodology has not been considered to propose.

Instead of this, the index structure and the obtained input variables are offered to the users, leaving them to determine the most suitable integration method according to the index destination and the availability of combination of its components with additional information that can be useful for the management.

In fact, a survey that was performed in Spain to potential users of this index (forest managers and technicians) in the framework of the FIREMAP research Project (http://www.geogra.uah.es/firemap) reveals the preference of managers and technicians to have the basic components of the index (variables and sub-indices), instead of the final index value.

Thus it enables the use of the information and its integration in the prevention systems available for each institution.

In order to make easier the integration of variables according to the user interests, the final risk index is generated from three sub-indices: the ignition danger index, the propagation danger index, and the vulnerability index.

These indices are obtained separately, in this way each one offers specific information about each need in terms of wildfire prevention (ignition, propagation, and wealth), and can be used independently for each management task.

The three indices combination will offer the global risk value, and depending on the user, the indices weighting can be varied according to the importance of each factor in the territory or in the considered management task.

It would be recommended for the integration of the variables that all indices that are taken into account for the global risk -not only the three intermediates, but also those that have been used to generate the three ones- were scaled in probability units (0-1 or 0-100%).

The way to achieve the conversion from the original index value to probability units has been included in this deliverable for some of the indices (i.e, fuel moisture content).

In any case, several straightforward solutions can be raised to make this conversion, using for instance linear functions from the maximum and minimum values adjustment.

The inclusion of scaled sub-indices to probability units will make easier the use of the different integration methods.

The hierarchical structure of the index, as well as the probability units format of the different sub-indices, will allow the inclusion of new variables or the substitution or removal of some of the proposed ones, in function of the available information of the territory in which will be used.

Moreover, it will allow introducing modifications in function of the work scale, as given some of the included variables have been presented in this deliverable (climate, population, etc.)

In the definition of the risk index structure, the proposed input variables have a spatial nature and a digital format, as far as possible according to the available information at the European Union level.

Remote sensing information (with a very suitable spatial and temporal component for this kind of studies) was considered, as well as point data (meteorological variables, census data, etc.) that would be spatialized by means of any spatial interpolation algorithm, and cartographic information available in digital format (Corine Land Cover maps, digital terrain models, national cartographic databases, etc.).

All this information leads to a suitable series of spatialized and geo-referenced information layers in digital format for using in a Geographical Information System (GIS)

In this context, GIS is a one of the most suitable tools for obtaining a risk wildfire index with the proposed characteristics in this deliverable, allowing the management, manipulation, analysis, modelling and representation of the input data as well as the output indices.

Moreover, they make considerably easier the updating processes of the information.

In this updating process, the existing connection between the GIS and the remotely sensed images processing should be considered, thus this is the source of an important part of the dynamic information used in the risk index (mainly fuel moisture content).

On the other hand, the disposal of this information in a GIS means that it can be used in forest management for different facets other than risk indices.

However, the use of this tool, that considerably facilitates the process, does not guarantees the goodness of the obtained index, which reliability will depend on the selection and accuracy of the input variables, as well as the success in the weighting risk assigned to each variable in the model.

Undoubtedly, GIS offer a wide range of possibilities to define the fire risk integration model:

EUFIRELAB

D-08-06.doc 101

- qualitative methods, where arbitrary weights are based on the judgment of an expert;

- quantitative indices, based on multi-criteria evaluation or other survey system;

- regression techniques, where statistical estimation methods are applied to explain fire occurrence;

- neural network, similar conceptually to the regression models, although with different fitting procedures. All these methods were analysed in detail in the deliverable D-08-05. Its use allows, with a good enough effective cost and speed, the creation of new models, the model application at different scenarios or the modification of risk classes limits, among others. Because of it, it seems unquestionable the use of this tool for the generation of the risk index proposed in this deliverable.

In these last years an important number of GIS applications have appeared in the web, with few use requisites, that make considerably easier the access to this application from everywhere.

The simplest ones only allow the visualization or download of a variable of interest, as the more complex ones allow the user to interact with the proposed models.

Several examples of both situations are found in the field of wildfire risk assessment.

For instance, in the United States we can obtain in real time the National Fire Danger Rating System estimation or any of its sub-indices that is composed (USDA Forest Service; http://www.fs.fed.us/land/wfas).

In the European Union, several risk indices by means of the Risk Forecast System can be obtained in the European Forest Fire Information System (EFFIS; http://effis.jrc.it/wmi/viewer.html).

In contrast to this simpler situation, in which the user only checks the generated information by the different institutions, situations that are more complex may appear, where the user can interact with the index generation.

One of the most interesting examples with this option is the Fire-Climate-Society Strategic Model (FCS-1) developed by the Arizona University (http://walter.arizona.edu),

The Fire-Climate-Society Strategic Model is part of the Model Wildfire Alternatives (WALTER), located at the University of Arizona.

WALTER is an interdisciplinary research initiative aimed at improving our understanding of the processes and consequences of interactions among wildfire, climate and society.

FCS-1 Fire-Climate-Society (FCS-1) is an online, spatially explicit strategic wildfire-planning model with an embedded multi-criteria decision process that facilitates the construction of user-designed risk assessment maps under alternative climate scenarios and varying perspectives of fire probability and values at risk.

The model is generated by integrating the following variables or sub-indices: fuel moisture stress index, fire return interval departure, large fire ignition probability, lightning probability, human factors of fire ignition, recreational value, species habitat richness, property value, personal landscape value.

One of the main advantages of this system is that it can be used with a standard model, where the weights of the different variables are assigned according to the model defined by the experts who created the system, or the user can decide the risk weights of each input variables, as shown in the figure 93.

This system seeks to capitalize on advances in geo-spatial, analytical, and web delivery technology to provide access to scientific and management activities.

Bearing in mind the required structure of the Euro-Mediterranean Wildland Fire Risk Index EM-WFRI: - the nature of variables that compose it, - the possibility of modifying and updating these

variables, - the integrating of different methods, - its use at different spatial scales, etc.,

it is obvious: - to recommend generating this risk index by means

of GIS and, - to integrate the use of this tool in a web service.

In this way a very flexible tool would be created, and could be used for wildfire management in very different facets of prevention and at different scales (local, regional, national, supranational), allowing the user to define the most suitable scenario at the time of use.

10.2 FIGURES

Figure 24: Analytic Hierarchy Process (AHP) weights to the Fire-Climate-Society Strategic Model

(Wildfire Alternatives, University of Arizona).

EUFIRELAB

D-08-06.doc 102

11 REFERENCES

ABHINEET J; RAVA SA; SINGH RK; DAS KK; ROY PS 1996. Forest fire risk modelling using remote sensing and geographic information system. Current Science (70 10):928-933.

AERIAL INFORMATION SYSTEMS Inc. 1981). Final report. Spatial risk method for predicting wildland fire occurrence. Crestline, California: USDA, Pacific Southwest Forest and Range Experiment Station.

AGUADO I; CHUVIECO E; CAMARASA A; MARTÍN MP; CAMIA A 1998). Estimation of Meteorological Fire Danger Indices from Multitemporal Series of NOAA-AVHRR data. Pages 1131-1147 in D. X. Viegas, editor. III International Conference on Forest Fire Research - 14th Conference on Fire and Forest Meteorology. ADAI, Coimbra.

AHLGREN IF; AHLGREN CE 1960. Ecological effects of forest fires. Bot. Rev. 26, 483–533.

ALBINI FA 1976. Estimating wildfire behavior and effects. USDA For. Serv. Gen. Tech. Rep. INT-30.

ALBINI FA; REINHARDT ED 1995. Modeling ignition and burning rate of large woody natural fuels. Int. J. Wildland Fire, 5: 81–91.

ALEXANDRIAN D; GOUIRAN M 1990. “Les causes d'incendies : levons le voile.” Revue forestière française - Espaces forestiers et incendies Numéro spécial 1990.

ALONSO M; CAMARASA A; CHUVIECO E; COCERO D; KYUN I; MARTÍN MP; SALAS FJ 1996. Estimating temporal dynamics of fuel moisture content of Mediterranean species from NOAA–AVHRR data. EARSEL Advances in Remote Sensing, 4(4), 9 – 24.

ALTOBELLIS A 1983. A survey of rural population density and forest fire occurrence in the South, 1956-1970. New Orleans: USDA Forest Service, Southern Forest Experimental Station. Report SO-294.

ANDERSEN HE; MCGAUGHLEY RJ; REUTEBUCH SE 2005. Estimating forest canopy fuel parameters using LIDAR data. Remote Sensing of Environment, 94, 441-449.

ANDERSON HE; SCHUETTE RD; MUTCH RW 1978. Timelag and equilibrium moisture content of ponderosa pine needles. Research Paper SE-74, USDA Forest Service, Intermountain Research Station, Ogden, Utah.

ANDREU V; RUBIO JL; FORTEZA J; CERNI R 1996. Postfire Effects on Soil Properties and Nutrient Losses. International Journal of Wildland Fire 6 (2), pp. 53-58.

ANDREU V; IMESON AC; RUBIO JL 2001. Temporal changes in soil aggregates and water erosion after a wildfire in a Mediterranean pine forest. Catena 44, 69-84.

BAILEY TC; GATRELL AC 1995. Interactive Spatial Data Analysis. Longman Scientific and Technical, Burnt Mill, Essex, England.

BAIRD AW. 1965. Attitudes and Characteristics of forest residents in three Mississippi counties., 48: Mississippi State University, Social Science Research Center Report No. 29.

BARBERIS A; DETTORI S; FILIGHEDDU MR. 2003. Management problems in Mediterranean cork oak forests: post-fire recovery. Journal of Arid Environments 54, 565-569.

BARET F.; FOURTY T. 1997. Estimation of leaf water content and specific leaf weight from reflectance and transmittance measurements. Agronomie, 17, 444- 455.

BARET F; GUYOT G. 1991. Potentials and limits of vegetation indices for LAI and APAR assessment. Remote Sensing of Environment, 35, 161-173.

BAYRAMIN I; ERPUL G; ERDOGAN EH. 2006. Use of CORINE methodology to Assess Soil Erosion Risk in the Semi-Arid Area of Beypazari, Ankara. Turk J Agric For, 30, 81-100.

BEMMERZOUK AM. 1997. Evaluation of the CFFDRS and the KBDI in the Mediterranean-type environment of Crete (Greece. M. Sc. Thesis, Chania, Greece.

BERRY J; BUCKLEY DJ; MCGARIGAL K. 1999. Integrating Landscape Structure Programs with GIS. www.innovativegis.com

BERTRAND AL; BAIRD AW. 1975. Incendiarims in Southern Forest: a decade of sociological research: Report. USDA Forest Service, Southern Forest Experimental Station and Social Science Research Center at Mississipi University.

BLACKMARR WH; FLANNER WB. 1968. Seasonal and diurnal variation in moisture content of six species of Pocosin shrubs. Research Paper: SE-33. pp. 11. USDA Forest Service, Southeastern Forest Experiment Station. Asheville, NC.

BLANCHI R ; JAPPIOT M et al. 2002. Forest fire assessment and cartography - a methodological approach. Forest fire research and wild land fire safety, Coimbra.

BOSSARD M; FERANEC J; OTAHEL J. 2000. CORINE land cover technical guide. Addendum 2000. European Environment Agency - Technical report n° 40.

BOVIO G; CAMIA A. 1998. An analysis of large forest fires danger conditions in Europe. Pages 975-994 in DX VIEGAS, editor. 3rd International Conference on Forest Fire Research, Luso-Coimbra, Portugal.

BOVIO G; CAMIA A. 2000. Description Eudic software for fire danger indices calculation.

EUFIRELAB

D-08-06.doc 103

BOVIO G; SOL B; VIEGAS DX. 1994. Synthese des Travaux sur l'intercomparison des indices de danger meteorologique d'incendie. Final Report of Minerve I Project EC, Environment and Climate Programme.

BOVIO G; CAMIA A.; GOTTERO F. 1999. Piano Regionale per la difesa del patrimonio boschivo dagli incendi 1999-2001. Regione Piemonte, Assessorato Economia Montana e Foreste.

BOWMAN WD 1989. The relationship between leaf water status, gas exchange, and spectral reflectance in cotton leaves. Remote Sensing of Environment, 30, 249– 255.

BRADSHAW LJ; DEEMING JE; BURGAN RE; COHEN JD 1983. The 1978 National Fire-Danger Rating System. Technical Documentation: GTR INT-169. pp. 44. USDA Forest Service. (Ogden, Utah.

BURGAN RE. 1979. Estimating live fuel moisture for the 1978. National Fire Danger Rating System. USDA For. Serv. Res. Pap. INT-226.

BURGAN RE; ROTHERMEL RC. 1984. BEHAVE: fire behaviour prediction and fuel modeling system. Fuel subsystem. USDA For. Serv. Gen. Tech. Rep. INT-167.

BURGAN RE; HARTFORD RA. 1993. Monitoring Vegetation Greenness with Satellite Data. GTR INT-297, USDA Forest Service, Ogden, Utah.

BURGAN, R. E., KLAVER, R. W. and KLAVER, J. M. 1998. Fuel models and fire potential from satellite and surface observations. International Journal of Wildland Fire 8:159-170.

BURGAN RE. 1999. Personal Communication. Not published.

CAMIA A; BOVIO G; AGUADO I; STACH N. 1999. Meteorological fire danger indices and remote sensing. Pages 39-59 in E. Chuvieco, editor. Remote Sensing of Large Wildfires in the European Mediterranean Basin. Springer-Verlag, Berlin.

CAMIA A ; BOVIO G ; MARZANO R. 2002. Preliminary analysis of wildland-urban interface fire prone areas in Italy. In: Viegas DX (ed.), Proc. IV Int. Conf. on Forest Fire Research and 2002 Wildland Fire Safety Summit, Luso (Portugal), Millpress Science Pub., Rotterdam, pp.9; ISBN 90-77017-72-0.

CAMIA A; VARELA V; MARZANO R; EFTICHIDIS G. 2004. Spatial analysis in European Wildland-Urban Interface environments using GIS. In: Xanthopoulos G. (ed.), Proc. II International Workshop “Forest Fire in the Wildland-Urban Interface and Rural Areas in Europe: an integral planning and management challenge”; 15–16 May 2003, Athens (Greece. Mediterranean Agronomic Institute of Chania (MAICh), Graphic Art and Publishing Unit, Chania, Crete, 185 – 196.

CARDILLE JA; VENTURA SJ; TURNER MG. 2001. Environmental and social factors influencing wildfires in the Upper Midwest, United States. Ecological Applications 11 (1):111-127.

CARLSON JD; BURGAN RE. 2003. Review of users’ needs in operational fire danger estimation: The Oklahoma example. International Journal of Remote Sensing, 24(8), 1601– 1620.

CARREGA P. 1990. Climatology and index of forest fire hazard in Mediterranean France. Pages B.05-01/11 in International Conference on Forest Fire Research, Coimbra, Portugal.

CARREGA P. 1994. Topoclimatologie et habitat. U.F.R. Espaces et Cultures. Nice, Université de Nice - Sophia Antipolis: 408.

CARTER GA. 1991. Primary and secondary effects of water content on the spectral reflectance of leaves. American Journal of Botany, 78, 916– 924.

CARVACHO L. 1998. Evaluación de la estimación de grandes incendios forestales en la cuenca mediterránea europea por redes neuronales y regresión logística. Serie Geográfica 7:73-85.

CASE P; BANKS B; BUTLER E; GOSNELL R. 2000. Chapter 7. Assessing Potential Wildfire Effects on People. Co-published simultaneously in Journal of Sustainable Forestry, 11(1/2), 159-175; and: Mapping Wildfire Hazards and Risks (ed: Neil Sampson, R. Dwight Atkinson, and Joe W. Lewis), Food Product Press, an imprint of The Haworth Press, Inc., 2000, , 159-175.

CASTRO FX; TUDELA A; SEBASTIÁ MA. 2003. Modeling moisture content in shrubs to predict fire risk in Catalonia (Spain. Agricultural and Forest Meteorology, 116, 49– 59.

CEC 1985. Soil Map of the European Communities, 1:1,000,000. CEC Luxembourg, 124pp., 7 maps.

CECCATO P; FLASSE S ; TARANTOLA S; JACQUEMOUD S; GRÉGOIRE JM. 2001. Detecting vegetation leaf water content using reflectance in the optical domain. Remote Sensing of Environment, 77, 22–33.

CECCATO P; GOBRON N; FLASSE S; PINTY B; TARANTOLA S. 2002. Designing a spectral index to estimate vegetation water content from remote sensing data: Part 1. Theoretical approach. Remote Sensing of Environment, 82, 188– 197.

CECCATO P; LEBLON B; CHUVIECO E; FLASSE S; CARLSON JD. 2003. Estimation of live fuel moisture content. In E. Chuvieco (Ed.), Wildland fire danger estimation and mapping. The role of remote sensing data (pp. 63–90. Singapore: World Scientific Publishing.

CERDÀ A; IMESON AC; CALVO A. 1995. Fire and aspect induced differences on the erodibility and hydrology of soils at La Costera, Valencia, south-east Spain. Catena 24, 289-304.

CHANDLER CP; CHENEY P; THOMAS P; TRABAUD L; WILLIAMS D. 1983. Fire in Forestry. Forest Fire behavior and effects. John Wiley and Sons: New York.

EUFIRELAB

D-08-06.doc 104

CHEN K; BLONG R; JAKOBSON C. 2003.Towards an Integrated Approach to Natural Hazards. Risk Assessment Using GIS: With Reference to Bushfires. Environmental Management, 31(4), 546-560.

CHLADIL MA; NUNEZ M. 1995. Assessing grassland moisture and biomass in Tasmania. The application of remote sensing and empirical models for a cloudy environment. International Journal of Wildland Fire, 5, 165– 171.

CHOU YH. 1992. Spatial autocorrelation and weighting functions in the distribution of wildland fires. International Journal of Wildland Fire 2 (24):169-176.

CHOU YH; MINNICH RA; CHASE RA. 1993. Mapping probability of fire occurrence in San Jacinto Mountains, California, USA. Environmental Management 17 (1):129-140.

CHRISTENSEN NL. 1987. The biogeochemical consequences of fire and their effects on the vegetation of the Coastal Plain of the southeastern United States. In: The role of fire in ecological systems, (L. Trabaud Ed.. SPB Academic Publishing 1- 19.

CHRISTIANSEN J; FOLKMAN W. 1971. Characteristics of people who start fires...some preliminary findings. Berkeley, Cal.: USDA, Forest Service. Pacific Southwest Forest and Range Experiment Station. Research Note PSW-251.

CHUVIECO E; CONGALTON RG. 1989. Application of remote sensing and Geographic Information Systems to Forest fire hazard mapping. Remote Sensing of Environment 29:147-159.

CHUVIECO E; SALAS FJ. 1996. Mapping the spatial distribution of forest fire danger using GIS. International Journal of Geographical Information Systems 10 (3):333-345.

CHUVIECO E; SALAS FJ; VEGA C. 1997. Remote sensing and GIS for long-term fire risk mapping. Pages 91-107 in E. Chuvieco, editor. A review of remote sensing methods for the study of large wildland fires. Departamento de Geografía, Universidad de Alcalá, Alcalá de Henares.

CHUVIECO E; DESHAYES M; STACH N; COCERO D; RIAÑO D. 1999a. Short-term fire risk: Foliage moisture content estimation from satellite data. In E. Chuvieco (Ed.), Remote sensing of large wildfires in the European Mediterranean Basin (pp. 17– 38. Berlin: Springer-Verlag.

CHUVIECO E; SALAS FJ; CARVACHO L; RODRÍGUEZ-SILVA F. 1999b. Integrated fire risk mapping. In Remote Sensing of Large Wildfires in the European Mediterranean Basin, Ed. E CHUVIECO. pp 61-84. Berlin: Springer-Verlag.

CHUVIECO E; RIAÑO D; AGUADO I; COCERO D. 2002. Estimation of fuel moisture content from multitemporal analysis of Landsat Thematic Mapper reflectance data: Applications in fire danger assessment. International Journal of Remote Sensing, 23(11), 2145–2162.

CHUVIECO E; AGUADO I; COCERO D; RIAÑO D. 2003a. Design of an empirical index to estimate fuel moisture content from NOAA–AVHRR analysis in forest fire danger studies. International Journal of Remote Sensing, 24(8), 1621– 1637.

CHUVIECO E; ALLGÖWER B; SALAS FJ. 2003b. Integration of physical and human factors in fire danger assessment. In Wildland Fire Danger Estimation and Mapping. The Role of Remote Sensing Data, ed. E. CHUVIECO, pp: 197-218. Singapore: World Scientific Publishing.

CHUVIECO E; AGUADO I; DIMITRAKOPOULOS A. 2004a. Conversion of fuel moisture content values to ignition potential for integrated fire danger assessment. Canadian Journal of Forest Research-Revue Canadienne De Recherche Forestiere 34 (11): 2284-2293.

CHUVIECO E; COCERO D; RIAÑO D; MARTÍN P; MARTÍNEZ-VEGA J., RIVA, J., and PÉREZ, F. 2004b. Combining NDVI and surface temperature for estimation of live fuel moisture content in forest fire danger rating. Remote Sensing of Environment, 92, 322-331.

CHUVIECO E; COCERO D; AGUADO I; PALACIOS-ORUETA A; Prado E. (in press. Improving burning efficiency estimates through satellite assessment of fuel moisture content. Journal of Geophysical Research - Atmospheres

COHEN WB. 1991. Response of vegetation indices to changes in three measures of leaf water stress. Photogrammetric Engineering and Remote Sensing, 57(2), 195– 202.

COLE LW; KAUFMAN HF. 1963. Socio-economic factors and forest fire in Mississippi counties. Mississippi. USA: Mississippi State University Social Sciences Research Centre. Preliminary Report 14.

COLL C; CASELLES V. 1997. A global split-window algorithm for land surface temperature from AVHRR data: Validation and algorithm comparison. Journal of Geophysical Research, 102B14, 16697–16713.

COLLINS DB. 1996. “FireLib User Manual and Technical Reference”. http://www.fire.org.

COMUNIDAD DE MADRID 2000): Plan de Protección Civil de emergencia por incendios forestales en la Comunidad de Madrid (INFOMA), BOCM nº 138 pp. 4-28.

CONSEIL DE L’EUROPE / COMMISSION DES COMMUNATES EUROPEENNES. 1987. Carte de la Végétation Naturelle des Etats membres des Communatés Européennes et du Conseil de l’Europe. Office for Official Publications of the European Communities, Luxembourg.

EUFIRELAB

D-08-06.doc 105

CORINE 1992. Soil Erosion Risk and Important Land Resources in the Southern Regions of the European Community. EUR 13233, Luxembourg.

CROSSON P 1997. Will erosion threaten agricultural productivity? Environment 39 (8), 4– 12.

CRUZ GM; BUTLER WB; ALEXANDER EM 2006. Predicting the ignition of crown fuels above a spreading fire. Part II: model evaluation. International Journal of Wildland Fire, Vol. 15, pp. 61-72.

CUTTER SL, MITCHELL JT, SCOTT MS 2000. Revealing the Vulnerability of People and Places: A Case Study of Georgetown County, South Carolina. Annals of the Association of American Geographers, 90(4), 713-737.

DATT B 1999. Remote sensing of water content in eucalyptus leaves. Australian Journal of Botany, 47, 909–923.

DE LA RIVA J , PEREZ F , RENAULT NL , KOUTSIAS N. 2004. Mapping forest fire occurrence at a regional scale. Remote Sensing of Environment 92(3):363-369

DE LUIS M, GONZALEZ-HIDALGO JC, RAVENTOS J. 2003. Effects of fire and torrential rainfall on erosion in a Mediterranean gorse community. Land Degradation and Development 14, 203-213.

DEBANO LF, KRAMMES JS 1966. Water repellent soils and their relation to wildfire temperatures. International Association of Hydrological Science Bulletin 2, 14-19.

DEBANO LF 1981. Water repellent soils: a state-of-the-art. USDA Forest Service General Technical Report PS W-46, Berkeley, CA.

DEBANO LF. 2000a. Water repellency in soils: a historical overview. Journal of Hydrology 231-231, 4-32.

DEBANO LF. 2000b. The role of fire and soil heating on water repellency in wildland environments: a review. Journal of Hydrology 231-231, 195-206.

DEEMING JE, BURGAN RE, COHEN JD. 1978. The National Fire-Danger Rating System - 1978. GTR INT-39, USDA Forest Service, Ogden, UT.

DEFRIES RS, TOWNSHEND JRG 1994. Global land-cover: Comparison of ground-based data sets to classifications with AVHRR Data. In G.M. Foody, and P.J. Curran (Eds.), Environmental remote sensing from regional to global scales ( pp. 84– 110. Chichester: John Wiley and Sons.

DE JONG SM, 1994. Applications of reflective remote sensing for land degradation Studies in a Mediterranean Environment. PhD thesis, Utrecht University, The Netherlands, 237 pp.

DELETIC A. 2001. Modelling of water and sediment transport over grassed areas. Journal of Hydrology 248, 168-182.

DELGADO XP 2001. Spatial patterns of forest fires in Catalonia (NE of Spain) along the period 1975-1995. Analysis of vegetation recovery after fire. Forest Ecology and Management, 147, pp.67-74 .

DELICHATSIOS MA, PANAGIOTOU T, KILEY F. 1991. The use of time-to-ignition data for characterizing the thermal inertia and the minimum (critical) heat flux for ignition or pyrolysis. Combust. Flame, 84: 323–332.

DENGIZ O, AKGUL S, 2005. Soil Erosion Risk Assessment of the Golbasi Environmental Protection Area and Its Vicinity Using the CORINE Model. Turk J Agric For, 29, 439-448.

DESBOIS N, DESHAYES M, BEUDOIN A. 1997. Protocol for fuel moisture content measurements. In 'A review of remote sensing methods for the study of large wildland fires. (Ed E. Chuvieco) pp. 61-72. Departamento de Geografía, Universidad de Alcalá. (Alcalá de Henares, Madrid.

DIMITRAKOPOULOS A, PAPAIOANNOU KK 2001. Flammability assessment of Mediterranean forest fuels. Fire Technology, 37, 143– 152.

DISASTER MONITORING CONSTELLATION http://directory.eoportal.org/pres_DMCDisasterMonitoringConstellationAlSAT1BILSAT1NigeriaSat1UKD

MCBeijing1.html

DOERR SH, SHAKESBY RA, WALSH RPD. 1998. Spatial variability of soil hydrophobicity in fire-prone eucalyptus and pine forests, Portugal. Soil Science 163, 313-323.

DOERR SH, SHAKESBY RA, WALSH RPD. 2000. Soil water repellency: its causes, characteristics and hydro-geomorphological significance. Earth-Science Reviews 51, 33-65.

DONOGHUE LR, MAIN WA. 1985. Some Factors Influencing Wildfire Occurrence and Measurement of Fire Prevention Effectiveness. Journal of Environmental Management 20 (1):87-96.

DONOGHUE LR, SIMARD AJ, MAIN WA. 1987. Determining the Economic Relationship Betwen Law Enforcement Activities and Arson Wildfires: a Feasibility Study in Arkansas. Journal of Environmental Management 25:377-393.

DOOLITTLE ML. 1979. Southern woods-burners: a descriptive analysis: USDA Forest Service. Southern Forest Experiment Station. Research Paper SO-151

DOOLITTLE ML, WELCH GD. 1974. Fire prevention in the Deep South: personal contact pays off. Journal of Forestry 72:488-490.

DOUGUEDROIT A, JOLY C. 2002. Spatial distribution of the forest fires in lower Provence : the influence of Marseille. Forest Fire Research and Wildland Fire safety, Coimbra, Milpress.

EUFIRELAB

D-08-06.doc 106

DRAKEA JB, DUBAYAHA RO, CLARK DB, KNOXD RG, BLAIRD BJ, HOFTONA MA, CHAZDONE RL, WEISHAMPELF JF, PRINCEA SD. 2002. Estimation of tropical forest structural characteristics using large-footprint lidar. Remote Sensing of Environment, 79, 305-319.

DROUET JC, SOL B. 1993. Mise au point d'un indice numerique de risque météorologique d'incendies de forêts. Forêt Mediterranéenne 14:155-162.

EASTMAN JR 1993. IDRISI. Clark University, Worcester.

ELVIDGE CD, BAUGH KB, KIHN E, KROEHL HW, DAVIS ER 1997. Mapping city lights with nightime data from the DMSP Operational Linescan System. Photogrammetric Engineering and Remote Sensing 63:727-734.

ESPELTA JM, RETANA J, HABROUK A 2003. An economic and ecological multi-criteria evaluation of reforestation methods to recover burned Pinus nigra forests in NE Spain. Forest Ecology and Management 180, 185-198.

EUROPEAN COMMISSION 2002. Forest Fires in Europe – 2001 fire campaign – Official Publication of the European Commission, S.P.I.02.72 EN.

EUROPEAN COMMISSION, JOINT RESEARCH CENTRE (DG JRC), Institute for Environment and Sustainability (IES), 2005) - CORINE Land Cover updating for the year 2000. IMAGE2000 and CLC2000. Products and Methods. Maria Vanda NUNES DE LIMA (Ed.. EUR 21757 EN, ISBN 92-894-9862-5.

EUROPEAN ENVIRONTMENTAL AGENCY 1996. Natural Resources, CD-Rom. European Environtmental Agency, Copenhagen.

EUROSTAT, 1999. Statistical references. OYce for oYcial publications of the European Communities, No.2/1999, Luxembourg.

FAO 1986. Wildland fire managemente terminology. Report nº 70, FAO Forestry Paper, Roma. M-99. ISBN 92-5-0024207.

F.A.O. 1988. Soil map of the world 1:500.000 scale. F.A.O. Rome.

FERREIRA AJD, COELHO COA, SHAKESBY RA, WALSH RPD 1997. Sediment and solute yield in forest ecosystems affected by fire and rip-ploughing techniques, central Portugal: a plot and catchment analysis approach. Physical Chemistry of the Earth 22, 309-314.

FOLKMAN WS. 1965. Forest fires as accidents: an epidemiological approach to fire prevention research, 15. Berkeley, California: Pacific Southwest Forest and Range Experiment Station. USDA.

FOLKMAN WS. 1973. Fire prevention in Butte County, California: their knowledge and attitudes regarding forest fire prevention., 32. Berkeley, California: USDA, Forest Service Pacific Southwest Forest Range Experiment Station. PSW-98.

FOSBERG MA, JE DEEMING. 1971. Derivation of the 1- and 10-hour timelag fuel moisture calculations for fire danger rating. Research Note RM-207, USDA, Forest Service, Fort Collins, Colorado.

FOSBERG MA, ROTHERMEL RC, ANDREWS PL. 1981. Moisture content calculations for 1000-Hour Timelag Fuels. Forest Science 27:19-26.

FOTHERINGHAM AS, BRUNSDON C, CHARLTON M. 2000. Quantitative Geography: Perspectives on Spatial Data Analysis. SAGE Publications, London.

FOTHERINGHAM AS, BRUNSDON C, CHARLTON M. 2002. Geographically weighted regression: The analysis of spatially varying relationships. John Wiley and Sons, New York, NY. 269 p.

FOX D, BEROLO W, CARREGA P, DARBOUX F 2006. Mapping erosion risk and selecting sites for simple erosion control measures after a forest fire in Mediterranean France. Earth Surface Processes and Landforms, Vol 31, pp. 606-621.

FUJIOKA FM. 1987. Fire weather network design: an update. In '9th Conference on fire and forest meteorology'. American Meteorological Society. pp. 165-169. (San Diego, California.

GAGARI P, LYRINTZIS G, BALOUTSOS G, XANTHOPOULOS G 2001. Contribution of log erosion barriers to soil protection and vegetation recovery after wildfire in a Pinus halepensis forest, in Greece. In Proceedings of the International Conference, Forest Research: a Challenge for an Integrated European Approach, Thessaloniki, Greece, Radoglou K. (ed.), 1:311-316.

GATRELL AC; BAILEY TC; DIGGLE PJ; ROWLINGSONT BS 1996. Spatial point pattern analysis and its application in geographical epidemiology. Transactions of the Institute of British Geographers 21:256-274.

GIOVANNINI G; S LUCHESI; M GIACHETTI; 1990. Beneficial and detrimental effects of heating on soil properties. In: Fire in ecosystem dynamics, (J.G. Goldamrner and M.J. Jenkins Eds. SPB Academic Publishing 95-102.

GIOVANNINI G, LUCCHESI S 1991. Is the vegetative cover the primary factor controlling erosion in burnt soils? In: Sala, M., Rubio, J.L. Eds. , Soil Erosion and Degradation as a Consequence of Forest Fires. Barcelona, p. 16.

GIOVANNINI G, VALLEJO R, LUCCHESI S, BAUTISTA S, CIOMPI S, LLOVET J. 2001. Effects of land use and eventual fire and soil erodibility in dry Mediterranean conditions. Forest Ecology and Management 147, 15-23.

GOBIN A, KIRKBY C, GOVERS G 2003. Pan-European soil erosion risk assessment. European Commision funded fifth framework project.

GOFAS A 2001. Εγχειρίδιο Δασοπροστασίας. Εκδόσεις Γιαχούδη-Γιαπούλη.

EUFIRELAB

D-08-06.doc 107

GONÇALVES MZ, LOURENÇO L. 1990. Meteorological index of forest fires risk in the Portuguese mainland territory. Pages B.07-01 - B.07-14 in I International Conference on Forest Fire Research., Coimbra.

GRANGER K. 1998. Developing an understanding of urban geohazard risk. Australian Journal of Emergency Management 13(4):13–17.

GULINCK H, MÚGICA M, DE LUCIO JV, ATAURI JA. 2001. A framework for comparative landscape analysis and evaluation based on land cover data, with an application in the Madrid region (Spain), Landscape and Urban Planning, 55 pp. 257-270.

HARDY CC. 2005. Wildland hazard and risk: problems, definitions and contexte. Forest ecology and management.

HARDY CC, BURGAN RE. 1999. Evaluation of NDVI for monitoring live moisture in three vegetation types of the Western U.S. Photogrammetric Engineering and Remote Sensing, 65, 603– 610.

HAYES JC, BARFIELD BJ, BARNHISEL RI 1984. Performance of grass filters under laboratory and field conditions. Transactions of the American Society of Agricultural Engineers 27, 1321-1331.

HONG S, BARRY H, DE ZONIA, MALDENOFF DJ. 2000. An aggregation index (AI) to quantify spatial patterns of landscape, Landscape Ecology, 15 pp. 591-601.

http://perso.wanadoo.fr/04as/carte/cartes/pacageol.htm

http://www.meteofrance.com/FR/climat/dpt_tempsdumois.jsp?LIEUID=DEPT06

http://www.prim.net/citoyen/definition_risque_majeur/introfeux.

http://www.promethee.com/prom/home.do (base de données Prométhée)

HUFFMAN EL, MCDONALD LH, STEDNICK JD. 2001. Strength and persistence of fire-induced soil hydrophobicity under ponderosa and lodgepole pine, Colorado front range. Hydrological Processes 15, 2877-2892.

ICONA. 1993. Manual de operaciones contra incendios forestales. 5.1/65, ICONA, Madrid.

IGIS 1997. FRAGSTATS*ARC User’s Manual www.innovativegis.com/products /fragstatsarc

ILLERA P, FERNANDEZ A, DELGADO JA. 1996. Temporal evolution of the NDVI as an indicator of forest fire danger. International Journal of Remote Sensing 17:1093-1105.

IMESON AC, VERSTRATEN JM, VAN MULLIGEN EJ, SEVINK J. 1992. The effects of fire and water repellency on infiltration and runoff under Mediterranean type forest. Catena 19, 345-361.

INBAR M, WITTENBERG L, TAMIR M. 1997. Soil Erosion and Forestry Management After Wildfire in a Mediterranean Woodland, Mt. Carmel, Israel. International Journal of Wildland Fire 7 (4), 285-294.

INBAR M, TAMIR M, WITTENBERG L. 1998. Runoff and erosion processes after a forest fire in Mount Carmel, a Mediterranean area. Geomorphology 24, 17-33.

JACKSON RD, IDSO SB, REGINATO RJ, PINTER PJ. 1981. Canopy temperature as a crop water stress indicator. Water Resources Research, 17, 1133–1138.

JACQUEMOUD S, USTIN SL, VERDEBOUT J, SCHMUCK G, ANDREOLI G, HOSGOOD B. 1996. Estimating leaf biochemistry using the PROSPECT leaf optical properties model. Remote Sensing of Environment, 56, 194–202.

JACQUEMOUD S, BACOUR C, POILVE H, FRANGI JP. 2000. Comparison of four radiative transfer models to simulate plant canopies reflectance: Direct and inverse mode. Remote Sensing of Environment, 74, 471– 481.

JOLY C, SUSINI F et al. 1999. Analysis of the occurrence of fire patterns in a limited area Forest fires: Needs and Innovations, Athens, Greece.

JONES AR, TAYLOR ML, BERTRAND AL. 1965. Some human factors in woods burning, 28. Baton Rouge, La.: La. State University Agricultural Experiment Station Bulletin Nº. 601.

KALABOKIDIS KD, KONSTANTINIDIS P, VASILAKOS C. 2002. GIS analysis of physical and impact on wildfire patterns. In Proceedings of IV International Conference on Forest fire research and wild land fire safety, Coimbra. (D.X. Viegas, Ed.)

KALLURI SNV, TOWNSHEND JRG, DORAISWAMY P. 1998. A simple single layer model to estimate transpiration from vegetation using multispectral and meteorological data. International Journal of Remote Sensing, 19(6), 1037– 1053.

KEETCH J, BYRAM G. 1968. A Drought Index for forest fire control. Research Paper. SE-38, USDA Forest Service, Southeastern Forest Experiment Station, Asheville, NC.

KERDILES H. 1997. Software for processing AVHRR data for the Communities of Europe (SAPCE): Algorithms used in SPACE version 2. Part 1: Calibration, atmospheric correction and cloud detection. Internal document at ARIS Unit, SAI, JRC Ispra.

KIRKBY M. 2001. MEDRAP Athens December 2001 Physical Aspects of Soil Erosion Control. page 11.

KOUTSIAS N, KALABOKIDIS K, ALLGÖWER B. 2002. Fire occurrence patterns at landscape level: beyond positional ac-curacy of ignition points with kernel density estimation methods. Poster presentation in 2002 World Conference on Natural Resources Modeling, “Modeling Natural and Biotic Resources in a Changing Planet”, 23-26 June 2002, Sigri, Lesbos.

EUFIRELAB

D-08-06.doc 108

KOUTSIAS N, KALABOKIDIS K, ALLGÖWER B. 2004. Fire occurrence patterns at landscape level: beyond positional accuracy of ignition points with kernel density estimation methods. Natural Resource Modeling 17(4): 359-375.

KOUTSIAS N, J MARTÍNEZ, E CHUVIECO, B ALLGÖWER. 2005. Modelling Wildland Fire Occurrence in Southern Europe by a Geographically Weighted Regression Approach. Paper presented at 5th International Workshop on Remote Sensing and GIS Applications to Forest Fire Management: Fire Effects Assessment, 16-18 June 2005, at Zaragoza, Spain.

KOUTSIAS N, MARTÍNEZ J, ALLGÖWER B. 2006. Global (OLS) and Local (Geographically--Weighted Regression) Models for Explaining Long-term Wildland Fire Occurrence Patterns in Southern Europe. Journal of Geophysical Research (in press)

KRAMMES JS, DEBANO LF. 1965. Soil Wettability: a neglected factor in watershed management. Water Resources Research 1, 283–286.

KUTIEL P, INBAR M. 1993. Fire impact on soil nutrients and soil erosion in a Mediterranean pine forest plantation. Catena 20, 129–139.

LAIRD JR, HARVEY MD. Complex response of a chaparral drainage basin to fire. In: Hadley, R.F. Ed. , Drainage Basin Sediment Delivery, Vol. 159. IAHS Publ., pp. 165–183.

LAL R, 2004. Soil carbon sequestration impacts on global climate change and food security. Science 304 (5677), 1623– 1627.

LE BISSONNAIS, Y. 1996. Aggregate stability and assessment of soil crustability and erodibility : I. Theory and methodology 47, 425-437.

LEBLON, B. 2001. Forest wildfire hazard monitoring using remote sensing: A review. Remote Sensing Reviews, 20(1), 1 –57.

LETEY, J. 2001. Causes and consequences of fire-induced soil water repellency. Hydrological Processes 15, 2867-2875.

LEVINE, N. 2002. CrimeStat II: A Spatial Statistics Program for the Analysis of Crime Incident Locations (version 2.0), Ned Levine and Associates Annandale, VA and The National Institute of Justice Washington, DC.

LICHSTEIN, J.W., SIMONS, T.R., SHRINER, S.A, FRANZREB, K.E. 2002. Spatial autocorrelation and autoregressive models in ecology. Ecological Monographs 72 (3): 445-463.

LIGDI, E., AND MORGAN, R.P.C. 1995. Contour grass strips: a laboratory simulation of their role in soil erosion control. Soil Technology 8, 109-117.

LOMBORG, B., 2001. The Skeptical Environmentalist: Measuring the Real State of the World. Cambridge University Press.

LOVELAND, T.R., MERCHANT, J.W., OHLEN, D.O. AND BROWN. J.F. 1991. Development of a Land-Cover Characteristics Database for the Conterminous U.S. Photogrammetric Engineering and Remote Sensing 57:1453-1463.

LYNHAM, T.J., AND MARTELL., D.L. 1985. An analysis of annual people-caused forest fire ocurrence in Ontario. Paper read at 8th National Conference on Fire and Forest Meteorology, April 29-May 2, at Detroit, Michigan.

MALDENOFF, D. J. AND BAKER, W.L. 2000): Spatial Modelling of Forest Landscape Change, Cambridge University Press.

MANDALLAZ, D., AND YE. R. 1997. Prediction of forest fires with Poisson models. Canadian Journal of Forest Research 27:1685-1694.

Manuel d’utilisation de SYSTAT 8.0 1998. Régression logistique : p 517 - 584

MARQUES, M.A., AND MORA E. 1998. Effects on erosion of two post-fire management practices: clear-cutting versus non-intervention. Soil and Tillage Research 45, 433-439.

MARTELL, D.L., OTUKOL, S. AND STOCKS, B.J. 1987. A logistic model for predicting daily people-caused forest fire occurrence in Ontario. Canadian Journal of Forest Research 17:394-401.

MARTELL, D.L., BEVILACQUA, E. AND STOCKS, B.J. 1989. Modelling seasonal variation in daily people-caused forest fire occurrence. Canadian Journal of Forest Research 19 (12):1555-1563.

MARTIN, C., CHEVALIER, Y., GIMENEZ, H., AND QUILLARD, J. 1998. Variations de la composition chimique des eaux de drainage et de ruissellement et modifications des caractères agronomiques des sols après incendie de forêt dans le bassin versant du Rimbaud (Var, France. Géomorphologie 3, 215-232.

MARTINEZ, J., CHUVIECO, E., MARTIN, M.P. 2004. Estimating human risk factors in Wildland fires in Spain using logistic regression. II International Symposium on Fire Economics, Planning and Policy: A Global Vision. Córdoba, Spain. 19-22 de Abril, 2004. 15 pp.

MARTINEZ, J, ROMERO, R. AND ECHEVARRIA, P. 2006): Valoración paisajística y ecológica de la Comunidad de Madrid: su integración en un índice sintético de riesgo de incendios forestales. XII Congreso Nacional de Tecnologías de la Información Geográfica. Granada. CDRom.

MATAIX-SOLERA, J., AND DOERR, S.H. 2004. Hydrophobicity and aggregate stability in calcareous topsoils from fire-affected pine forests in south-eastern Spain. Geoderma 118, 77-88.

MCARTHUR, A.G. 1967. Fire behaviour in eucalypt forests. Leaflet 107, Commonwealth of Australia Forest and Timber Bureau, Canberra, Australian Capital Territory.

EUFIRELAB

D-08-06.doc 109

MCGARIGAL, K. AND MARKS, B.J. 1995): FRAGSTATS: spatial pattern analysis program for quantifying landscape structure. Portland, Department of Agriculture, Forest Service.

MCGARIGAL, K., MARKS, B., HOLMES, C. AND ENE, E. 2002): Fragstats 3.3. Spatial Pattern Analysis Program for Quantifying Landscape Structure, Department of Natural Resources Conservation, University of Massachusetts.

MONTANARELLA, L., 2003. Soil erosion risk in Europe. European Commision, Joint Research Centre, Institute for Environment and Sustainability. T.P. 280, 21020 Ispra (VA) ITALY. http://ies.jrc.cec.eu.int/Projects/ESB/.

MORAN, M.S., CLARKE, T.R., INOUE, Y., AND VIDAL, A. 1994. Estimating crop water deficit using the relation between surface-air temperature and spectral vegetation index. Remote Sensing of Environment, 49, 246– 263.

MORGAN, P., HARDY, C.C., SWETNAM, T.W., ROLLINS, M.G., AND LONG, D.G. 2001. Mapping fire regimes across time and space: Understanding coarse and fine-scale fire patterns. International Journal of Wildland Fire, Vol. 10: 329-342.

MORSDORF, F., MEIER, E., KOTZ, B., ITTEN, K. I., DOBBERTIN, M. ALLGOWER, B. 2004. LIDAR-based geometric reconstruction of boreal type forest stands at single tree level for forest and wildland fire management. Remote Sensing of Environment, 92, 353-362.

MUNOZ-CARPENA, R., PARSONS, J.E., AND GILLIAM, J.W. 1999. Modeling hydrology and sediment transport in vegetative filter strips. Journal of Hydrology 214, 111-129.

NAVEH, Z. AND LIEBERMAN, A. 1994): Landscape Ecology. Theory and Application, Nueva York, Springer-Verlag.

NEARING, M.A., PRUSKI, F.F., O’NEAL, M.R., 2004. Expected climate change impacts on soil erosion rates: a review. Journal of Soil and Water Conservation 59 (1), 43– 50.

NEARY, D.G., KLOPATEK, C.C., DEBANO, L.F., AND FOLLIOTT, P.F. 1999. Fire effects on belowground sustainability: a review and synthesis. Forest Ecology and Management 122, 51-71.

NEARY, G. D., 2004. An Overview of Fire Effects on Soils. Southwest Hydrology, September/October 2004.

NEUENSCHWANDER, L.F., MENAKIS J.P., MILLER, M., SAMPSON, R.N., HARDY, C., AVERILL, B., MASK, R., 2000. Chapter 3. Indexing Colorado Watersheds to Risk of Wildfire. Co-published simultaneously in Journal of Sustainable Forestry, 11(1/2), 159-175; and: Mapping Wildfire Hazards and Risks (ed: Neil Sampson, R. Dwight Atkinson, and Joe W. Lewis), Food Product Press, an imprint of The Haworth Press, Inc., 2000, , 35-55.

NOBLE, I.R., BARY, G.A.V. AND GILL. A.M. 1980. McArthur's fire-danger meters expressed as equations. Australian Journal of Ecology 5:201-203.

ORIEUX, A. 1979. Conditions météorologiques et incendies di forêts en région mèditerranéenne. Note Technique du Service Météorologie Métropolitain Ministère des Transports, Direction de la Météorologie.

PALMIERI, S., INGHILESI, R., SIANI, A.M. AND MARTELLACCI., C. 1992. Un indice meteorologico di rischio per incendi boschivi. Bollettino Geofisico 15:49-62.

PALTRIDGE, G.W., AND BARBER, J. 1988. Monitoring grassland dryness and fire potential in Australia with NOAA/AVHRR data. Remote Sensing of Environment, 25, 381– 394.

PAPAMIXOS, T., N., 1996. Δασικά Εδάφη, Σχηματισμός, Ιδιότητες, Συμπεριφορά. Aristotle University of Thessaloniki, Υπηρεσία Δημοσιευμάτων.

PAUSAS, J.G. 1997. Resprouting of Quercus suber in NE Spain after fire. Journal of Vegetation Science 8, 703-706.

PAUSAS, J.G. 2004. Changes in fire and climate in the eastern Iberian peninsula (Mediterranean basin. Climatic Change 63, 337-350.

PAUSAS, J.G, AND VALLEJO, V.R. 1999. The role of fire in European Mediterranean Ecosystems. In Remote Sensing of Large Wildfires in the European Mediterranean Basin, Chuvieco E. (ed.), Springer-Verlag, 3-16.

PEW, K.L. AND LARSEN, C. 2001. “GIS analysis of spatial and temporal patterns of human-caused wildfires in the temperate rain forest of Vancouver island, Canada.” Forest ecology and management 140: 1-18

PHILLIPS, C.B., AND NICKEY. 1978. The concept of "spatial risk" and its application to fire prevention. Fire Management Notes 39 (4):7-8-19.

PINAYA, I, SOTO B, ARIAS M, AND DÍAZ-FIERROS F. 2000. Revegetation of burnt areas: relative effectiveness of native and commercial seed mixtures. Land Degradation and Development 11, 93-98.

PROSPER-LAGET, V., DOUGUÉDROIT, A., AND GUINOT, J.P. 1995. Mapping the risk of forest fire occurrence using NOAA satellite information. EARSEL Advances in Remote Sensing, 4(3), 30– 38.

PROSSER IP, AND WILLIAMS L. 1998. The effect of wildfire on runoff and erosion in native Eucalyptus forest. Hydrological Processes 12, 251-265.

RENARD, K.G., FOSTER, G.R., WEESIES, G.A., MCCOOL, D.K., YODER, D.C., 1997. Predicting soil erosion by water: a guide to conservation planning with the Revised Universal Soil Loss Equation. Agricultural Handbook, vol. 703. U.S. Department of Agriculture 404 pp.

EUFIRELAB

D-08-06.doc 110

RIAÑO, D., MEIER, E., ALLGÖWER, B., CHUVIECO, E. AND USTIN, S. L. 2003. Modelling airborne laser scanning data for the spatial generation of critical forest parameters in fire behavior modelling. Remote Sensing of Environment, 86, 177– 186.

RICHARD, J.-F., AND BEROUTCHACHVILI, N. 1996. “Vers l'élaboration d'un système d'information sur les paysages du monde.” Cahiers des sciences humaines 32: 823-842.

RIITTERS, K.H; O’NEILLS, R.V; HUNSAKER, C.T; WICKHAM, J.D; YANKEE, D.H. AND TIMMINS, S.P. 1995): A factor analysis of landscape pattern and structure metrics, Landscape Ecology, 10, pp. 23-39.

RIVA, J.D.L., PEREZ-CABELLO, F. et al. 2004. “Mapping Wildfire occurrence at regional scale.” Remote sensing of environnement 92: 363-369.

ROBICHAUD, PR, HUNGERFORD RD. 2000. Water repellency by laboratory burning of four northern Rocky Mountain forest soils. Journal of Hydrology 231–232: 207–219.

ROBICHAUD, P. 2004. Post-fire rehabilitation treatments: are we learning what works ? Southwest hydrology, Watersheds on Fire 3, 20-21.

RODRIGUEZ Y SILVA, F. 2002. Visual Forest Fire Weather Index (vffwi) a mathematical model for the prediction of forest fires weather danger in mediterranean ecosystems. In 'IV Conference on Forest Fire Research'. ADAI, (Luso,Portugal.

ROTHERMEL, R.C. 1972. A Mathematical Model for Predicting Fire Spread in Wildland Fuels. Res. Pap. INT-115. pp. 40. USDA Forest Service. (Ogden, Utah.

ROTHERMEL, R.C., WILSON, R.A. MORRIS, G. A., AND SACKETT., S.S. 1986. Modelling moisture content of fine dead wildland fuels: input to BEHAVE fire prediction system. Research Paper INT-359, USDA Forest Service Intermountain Research Station, Odgen, Utah.

RUBIO, J.L, FORTEZA, J., ANDREU, V., AND CERNI, R. 1997. Soil profile characteristics influencing runoff and soil erosion after forest fire: a case study (Valencia, Spain. Soil Technology 11, 67-78.

RUBIO, J. L., BOARDMAN J., HELMING K., 2006. Soil erosion across Europe: Research approaches and perspectives. Catena, Article in Press.

SADAHIRO, Y. 2003. Stability of the surface generated from distributed points of uncertain location. International Journal of Geographical Information Science 17:139-156.

SALAS, F.J., AND CHUVIECO, E. 1994. GIS applications to forest fire risk mapping. Wildfire 3:7-13.

SAMPSON, R.N., NEUENSCHWANDER L.F., 2000. Chapter 2. Characteristics of the Study Area, Data Utilized, and Modeling Approach. Co-published simultaneously in Journal of Sustainable Forestry, 11(1/2), 159-175; and: Mapping Wildfire Hazards and Risks (ed: Neil Sampson, R. Dwight Atkinson, and Joe W. Lewis), Food Product Press, an imprint of The Haworth Press, Inc., 2000, , 15-33.

SAN MIGUEL-AYANZ, J., ANNONI, A. AND SCHMUCK. G. 1998. The use of satellite imagery for retrieval of information on wildfire damage in Mediterranean landscapes. in ERIM’98, International Symposium on Remote Sensing of Environment: Information for sustainability, Tromso.

SAN MIGUEL-AYANZ, J., BARBOSA, P., SCHMUCK, G., AND LIBERTA, G. 2002. Towards a coherent forest fire information system in Europe: The European Forest Fire Information System (EFFIS. in D. X. Viegas, editor. IV International Conference on Forest Fire Research and 2002 Wildland Fire Safety Summit. ADAI, University of Coimbra, Coimbra, Portugal.

SANTOS, MÁRIO R. CAETANO, PAULO M. BARBOSA AND JOSÉ U. PAÚL, 2000. A comparative study of vegetation indices to assess land cover change after forest fires. Remote Sensing for Earth Science, Ocean, and Sea Ice Applications, 3868, pp.232-240.

SCOTT, D.F., VAN WYK, D.B., 1990. The effects of wildfire on soil wettability and hydrological behaviour of an afforested catchment. Journal of Hydrology 121, 239–256.

SEBASTIÁN LÓPEZ, A. 2004. Modelización a escala europea del riesgo de incendio forestal, desde la perspectiva dinámica y estructural. Universidad de Valladolid, Valladolid.

SEBASTIÁN LÓPEZ, A., SALVADOR, R., AND SAN-MIGUEL-AYANZ, J. 2001. Two models for a structural forest fire risk index at a regional scale. in 21st EARSeL Workshop Remote sensing and GIS applications to forest fire management: New Methods and Sensor, Paris. France.

SEBASTIÁN LÓPEZ, A., SAN-MIGUEL-AYANZ, J., AND BURGAN, R.E. 2002. Integration of satellite sensor data, fuel type maps and meteorological observation for evaluation of forest fire risk at the pan-European scale. International Journal of Remote Sensing 23(13): 2713-2719.

SEVINK, J., IMESON, A.C., VERSTRATEN, J.M., 1989. Humus form development and hillslope runoff, and the effects of fire and management under mediterranean forest in NE-Spain. Catena 16, 461–475.

SHAKESBY, R.A., COELHO, C.O.A., FERREIRA, A.J.D., TERRY, J.P., and WALSH, R.P.D. 1993. Wildfire impacts on soil erosion and hydrology in wet Mediterranean forest, Portugal. International Journal of Wildland Fire 3, 95-110.

EUFIRELAB

D-08-06.doc 111

SHAKESBY, R.A., BOAKES, D.J., COELHO, C.O.A., GONÇALVES, A.J.B., AND WALSH, R.P.D. 1994. Limiting soil loss after forest fire in Portugal: the influence of different post-fire timber clearing practices. In Proc. 2nd Int. Conf. Forest Fire Research II, 1161-1170.

SHAKESBY, R.A., BOAKES, D.J., COELHO, C.O.A., GONÇALVES, A.J.B., AND WALSH, R.P.D. 1996. Limiting the soil degradational impacts of wildfire in pine and eucalyptus forests in Portugal. Applied Geography 16, 337-355.

SHAKESBY, R.A, DOERR, S.H, AND WALSH, R.P.D. 2000. The erosional impact of soil hydrophobicity: current problems and future research directions. Journal of Hydrology 231-232, 178-191.

SILVA, F.C.O.R. 2002. "Visual forest fire weather index (VFFWI)" a mathematical model for the prediction of forest fire weather danger in Mediterranean ecosystems. Forest fire research and wild land fire safety, Coimbra.

SILVERMAN, B.W. 1986. Density Estimation for Statistics and Data Analysis. Chapman and Hall, London, England.

SIMARD, A.J. 1968. The moisture content of forest fuels. A review of the basic concepts. Information Report: FF-X-14. pp. 47. Forest Fire Research Institute. (Ottawa, Ontario.

SMITH, S.J., WILLIAMS, J.R., MENZEL, R.G., COLEMAN, G.A., 1984. Prediction of sediment yield from Southern Plains grasslands with the Modified Universal Soil Loss Equation. Journal of Range Management 37 (4), 295– 297.

SNEEUWJAGT, R.J., AND PEET, G.B. 1985. Forest fire behaviour tables for Western Australia. Perth: Western Australian Department of Conservation and Land Management.

SOL, B. 1990. Estimation du Risque Météorologique D'incendies de Forêts dans le Sud-Est de la France. Revue Forestière Française 17:263-271.

SOL, B. 1999. A European program improves the operational fight against forest fires: MINERVE Program and meteorological fire danger in southeastern France, in DELFI Proceedings, Athens, pp. 185-192.

SOLER, M., SALA, M., 1992. Effects of fire and of clearing in a Mediterranean Quercus ilex woodland: an experimental approach. Catena 19 3r4 , 321–332.

SOTO, B., BENITO, E., BASANTA, R., DIAZ-FIERROS, F., 1991. Influence of antecedent soil moisture on pedological effects of fire. In: Sala, M., Rubio, J.L. Eds., Soil Erosion and Degradation as a Consequence of Forest Fires. Barcelona.

SOTO, B. and DÍAZ-FIERROS, F. 1998. Runoff and soil erosion from areas of burnt scrub: comparison of experimental results with those predicted by the WEPP model. Catena 31, 257-270.

SOTO, B, BASANTA R, PEREZ R, AND DIAZ-FIERROS F. 1995. An experimental study of the influence of traditional slash-and-burn practices on soil erosion. Catena 24, 13-23.

STRAHLER, A.H., WOODCOCK, C.E., AND SMITH, J.A. 1986. On the nature of models in remote sensing. Remote Sensing of Environment, 20, 121–140.

SWANSON, F.J., 1981. Fires and geomorphic processes. In: Proceedings, fire regimes and ecosystems conference, Honolulu. Gen. Tech. Rep. WO-26 USDA, Washington, DC, pp. 401–420.

TADESSE, L.D. and MORGAN, R.P.C., 1996. Contour grass strips: a laboratory simulation of their role in erosion control using live grasses. Soil Technology 9, 83-89.

TERRES, J.M. 1999. The Crop Growth Monitoring System implemented by JRC/ARIS unit for the information needs of the EC DG VI - Agriculture. Proceedings of the international symposium Modelling Cropping Systems - European Society for Agronomy. EX 18746/C. ORA/PRO 60194R

THOMAS, A.D., WALSH, R.P.D, AND SHAKESBY, R.A. 1999. Nutrient losses in eroded sediment after fire in eucalyptus and pine forests in the wet Mediterranean environment of northern Portugal. Catena 36, 283-302.

THOMAS, A.D., WALSH, R.P.D., AND SHAKESBY, R.A. 2000. Solutes in overland flow following fire in eucalyptus and pine forests, northern Portugal. Hydrological Processes 14, 971-985.

THOMPSON, W.A. 2000. Using forest fire hazard modelling in multiple use forest management planning. Forest Ecology and Management 134:163-176.

THORNES, J.B., 1985. The ecology of erosion. Geography 70, 222–234.

TRABAUD, L. 1976. Inflamabilité et combustibilité des principales espèces des garrigues de la region mediterranéenne. Oecol. Plant. 11: 117–136.

TUCKER, C.J., J.R. TOWNSHEND AND T.E. GOFF. 1985. African land-cover classification using satellite data. Science, 227: 367-374.

VACCA, A., LODDO, S., OLLESCH, G., PUDDU, R., SERRA, G., Tomasi, D. and Aru, A. 2000. Measurement of runoff and soil erosion in three areas under different land use in Sardinia (Italy. Catena 40, 69-92.

VALETTE, J.C. 1990. “Inflammabilité des espèce forestières méditerranéennes: conséquences sur la combustibilité des formations forestières.” Revue forestière française - espaces forestiers et incendies numéro spécial 1990.

EUFIRELAB

D-08-06.doc 112

VALETTE, J.C. 1992. Inflammabilities of Mediterranean species. In Forest Fire Risk and Management: European School of Climatology and Natural Hazards Course. Halkidiki, Greece, 27 May – 4 June 1992. Edited by P. Balabanis, G. Eftichidis, and R. Fantechi. European Commission, Halkidiki, Greece. pp. 51–64.

VALLEJO, R, AND ALLOZA J.A. 1998. The restoration of burned lands: the case of eastern Spain. Large Forest Fires, 91-108.

VAN DER GOOT, E, 1997. Technical description of interpolation and processing of meteorological data in CGMS. Internal document at SAI, JRC Ispra

VAN WAGNER, C.E. 1982. Initial moisture content and the exponential drying process. Canadian Journal of Forest Research 12:90-92.

VAN WAGNER, C.E. 1987. Development and structure of the Canadian Forest Fire Weather Index System. Technical Report 35. pp. 37. Canadian Forestry Service. (Ottawa, Ontario.

VASCONCELOS, M.J.P., SILVA, S., TOMÉ, M., ALVIM, M., and J.M.C. PEREIRA. 2001. Spatial prediction of fire ignition probabilities: comparing logistic regression and neural networks. Photogrammetric Engineering and Remote Sensing 67 (1):73-83.

VEGA GARCÍA, C., WOODARD, P.M. TITUS, S.J. ADAMOWICZ, W.L. AND LEE, B.S. 1995. A logit model for predicting the daily occurrence of human caused forest fires. International Journal of Wildland Fire 5 (2):101-111.

VEGA GARCÍA, C., LEE, B.S., WOODARD, P.M., AND TITUS. 1996. Applying neural network technology to human-caused wildfire occurrence prediction. AI Applications 10 (3):9-18.

VELEZ, R et al. 2000 La defensa contra incendios forestales: fundamentos y experiencias. McGraw Hill, Madrid

VELEZ, R. AND MÉRIDA, J.C. 2002. Forest fire management in Spain: Some examples of systematic analysis of a comprehensive database to improve effectiveness and efficiency. Forest fire research and Wild land fire safety, Coimbra.

VELEZ, R., REANO, P., et al. 2002. Forest fire prevention with a target: the rural people. Forest fire research and wildland fire safety, Coimbra.

VERSTRAETE, M.M. AND PINTY, B. 1996. Designing optimal spectral indexes for remote sensing applications. IEEE Transactions on Geoscience and Remote Sensing, 34(5), 1254– 1265.

VIDAL, A., PINGLO, F., DURAND, H., DEVAUX-ROS, C., AND MAILLET, A. 1994. Evaluation of a temporal fire risk index in Mediterranean forest from NOAA thermal IR. Remote Sensing of Environment, 49, 296– 303.

VIEDMA, O., MELIA, A, J., SEGARRA, D., GARCÍA-HARO, J., 1997. Modelling rates of ecosystem recovery after fires by using Landsat TM data. Remote Sensing of Environment, 61(3), pp. 383-398.

VIEGAS, D.X., ed. 1998. Proceedings of the III International Conference on Forest Fire Research and 14th Conference on Fire and Forest Meteorology. Held in Luso, Coimbra, Portugal, on November 16-20th: ADAI - Associaçao para o Desenvolvimento da aerodinâmica Industrial - and University of Coimbra.

VIEGAS, D.X. 1998b. Fuel moisture evaluation for fire behaviour assessment. In G. Eftichidis, P. Balabaris, and A. Ghazi (Eds.), Advanced study course on wildfire management. Final Report, Marathon (pp. 81– 92. Athens: European Union.

VIEGAS, D.X., ed. 2002. Proceedings of the IV International Conference on Forest Fire Research and 2002 Wildland Fire Safety Summit. Held in Coimbra, Portugal, on 18-23 November: ADAI - Associaçao para o Desenvolvimento da aerodinâmica Industrial - and University of Coimbra, Portugal.

VIEGAS, D.X., VIEGAS, T.P., AND FERREIRA, A.D. 1992. Moisture content of fine forest fuels and fire occurrence in central Portugal. International Journal of Wildland Fire, 2(2), 69–85.

VIEGAS, D.X., BOVIO, G., CAMIA, A., FERREIRA, A., AND SOL. B. 1996. Testing Meteorological Fire Danger Methods in Southern Europe. Pages 571-589 in 13th Conference on Fire and Forest Meteorology, Lorne - Australia.

VIEGAS, D.X., PIÑOL, J., VIEGAS, M.T., AND OGAYA, R. 2001. Estimating live fine fuels moisture content using meteorologically-based indices. International Journal of Wildland Fire, 10, 223–240.

VINEY, N.R. 1991. A review of fine fuel moisture modelling. International Journal of Wildland Fire 1(4): 215-234.

VOUGHT, L.B.M, PINAY, G., FUGLSANG, A., AND RUFFINONI, C. 1995. Structure and function of buffer strips from a water quality perspective in agricultural landscapes. Landscape and Urban Planning 31, 323-331.

WALLIS, M.G., AND HORNE, D.J. 1992. Soil water repellency. In Advances in Soil Science, Vol. 20, Stewart BA (ed., Springer: New York, 91-146.

WALSH, R.P.D., COELHO, C.O.A., SHAKESBY, R.A., AND TERRY, J.P. 1992. Effects of land use management practices and fire on soil erosion and water quality in the Agueda river basin, Portugal. Geoökoplus III, 15-36.

EUFIRELAB

D-08-06.doc 113

WALSH, R.P.D., BOAKES, D.J., COELHO, C.O.A., FERREIRA, A.J.D., SHAKESBY, R.A., AND THOMAS, A.D. 1995. Post-fire land use and management and runoff responses to rainstorms in northern Portugal. In Geomorphology and Land Management in a Changing Environment, McGregor DFM, and Thompson DA (eds), Wiley,Chichester, 283-308.

WARREN, A., 2002. Land degradation is contextual. Land Degradation & Development 13 (6), 449– 459.

WOHLGEMUTH, P.M., HUBBERT, K.R., AND ROBICHAUD, P.R. 2001. The effects of log erosion barriers on post-fire hydrologic response and sediment yield in small forested watersheds, southern California. Hydrological Processes 15, 3053-3066.

WORTON, B.J. 1989. Kernel methods for estimating the utilization distribution in homerange studies. Ecology 70:164-168.

YOOL, S.R., ECKHARDT, D.W., ESTES, J.E., AND COSENTINO., M.J. 1985. Describing the brushfire hazard in southern California. Annals of the Association of American Geographers 75 (3):417-430.

YOUNG, R.A., ONSTAD, C.A., BOSCH, D.D., ANDERSON, W., 1989. AGNPS: a nonpoint-source pollution model for evaluating agricultural watersheds. Journal of Soil and Water Conservation 44 (2), 168–173.

ZARCO-TEJADA, P.J., RUEDA, C.A., AND USTIN, S.L. 2003. Water content estimation in vegetation with MODIS reflectance data and model inversion methods. Remote Sensing of Environment, 85, 109– 124.

ZHANG, L.J., SHI, H.J. 2004. Local modeling of tree growth by geographically weighted regression. Forest Science 50 (2): 225-244.

ZHANG QUANFA, GORAN PAVLIC, WENJUN CHEN, RASIM LATIFOVIC, ROBERT FRASER, JOSEF CIHLAR, 2004. Deriving stand age distribution in boreal forests using SPOT VEGETATION and NOAA AVHRR imagery. Remote Sensing of Environment, 91, pp. 405–418.

EUFIRELAB

D-08-06.doc 114

12 ANNEX: MAPPING POST-FIRE SOIL EROSION RISK

FOX D, BEROLO W, CARREGA P

UMR 6012 Espace CNRS, Equipe GVE, Department of Geography, University of Nice Sophia Antipolis BP 3209 06204 Nice Cedex 3, France.

12.1 INTRODUCTION

The Mediterranean climate is particularly conducive to large-scale forest fires.

Mean annual precipitation is sufficient to support a dense mixture of oak, pine, and other forest species and the warm dry summer season makes both the live and accumulated biomass highly inflammable for up to at least two months of the year (VALLEJO and ALLOZA, 1998; PAUSAS and VALLEJO, 1999.

The abandonment of traditional underbrush clearing practices combined with an increase in accidental and intentional fire ignitions has led to more frequent forest fires throughout much of the Mediterranean region (MORENO and VALLEJO, 1999, in FERNANDEZ et al., 2003).

Changing climatic conditions are likely to increase the number of fires and area burned in coming years (PAUSAS, 2004).

The impacts of a major forest fire on runoff and erosion are well-known, and these include greater peak flows and soil loss until the return of a vegetation cover.

The first winter following a summer forest fire is the most vulnerable period and runoff and erosion rates generally decrease rapidly afterwards (SOTO and DÍAZ-FIERROS, 1998; VACCA et al., 2000).

It is therefore essential that runoff and erosion management strategies be implemented in the first weeks after a fire.

Several natural factors determine the impact of a fire on post-fire runoff and erosion rates, and these include the following: pre-fire vegetation, topography, slope aspect, fire severity, changes in soil properties, and post-fire rainfall (WALSH et al., 1992; RUBIO et al., 1997; INBAR et al., 1998; SOTO and DÍAZ-FIERROS, 1998; THOMAS et al., 2000; DELUIS et al., 2003).

Although many of these factors (such as slope or rainfall intensity) are common to all soil erosion contexts and need little explanation, the changes in soil erodibility brought about by the intense heat are particular to forest fire conditions.

For this reason, these will be reviewed briefly before presenting the erosion control strategy.

12.2 THE IMPACT OF FOREST FIRES ON SOIL ERODIBILITY

The increases in runoff and soil erosion rates observed after a forest fire are due primarily to the destruction of the vegetation cover (PROSSER and WILLIAMS, 1998; WOHLGEMUTH et al., 2001).

However, the combustion of the standing vegetation and litter layer provokes several changes in soil properties.

The study of low severity fires or traditional slash and burn methods has confirmed some of the short-term benefits of forest fires on soil fertility (pH, nutrient status) and their influence on post-fire regrowth (SOTO et al., 1995; NEARY et al., 1999).

However, the effects are not all beneficial since the combustion of the soil organic matter can also lead to lower cation exchange capacities, and high severity fires can lead to significant nutrient losses through volatilisation and accelerated erosion (MARTIN et al., 1998; NEARY et al., 1999; THOMAS et al., 1999).

Two further changes in soil characteristics, the development or enhancement of hydrophobicity and a decrease in aggregate stability, have a direct impact on soil erodibility (ANDREU et al., 2001.

The formation of soil hydrophobicity, or water repellency, has been described extensively elsewhere (e.g. WALLIS and HORNE, 1992; DEBANO, 2000; DOERR et al., 2000; DEBANO, 2000b; HUFFMAN et al., 2001), so only a few aspects related directly to runoff generation and soil erosion will be considered here.

Water repellency develops naturally in most Mediterranean forests (DOERR et al., 1998; MATAIX-SOLERA and DOERR, 2004).

During a forest fire, heat from the combustion vaporizes organic substances, some of which migrate downwards into the soil where they condense at cooler temperatures and coat mineral particles (DEBANO, 2000)

After a forest fire, it is therefore common to find a thin layer of hydrophobic soil at the surface or within a depth of a few centimetres (DEBANO, 2000; MATAIX-SOLERA and DOERR, 2004).

Although the effect of hydrophobicity on runoff and erosion has clearly been demonstrated at the plot scale, it has been difficult to demonstrate its importance for a catchment (SHAKESBY et al., 1993; SHAKESBY et al., 2000).

One reason for this is the high spatial variability in hydrophobicity: at the catchment scale, preferential flow in decayed root channels, cracks, rodent burrows, and hydrophilic patches may account for a large proportion of the infiltrated water (IMESON et al., 1992; FERREIRA et al., 1997; SHAKESBY et al., 2000).

EUFIRELAB

D-08-06.doc 115

The impact of forest fires on soil aggregate stability is more ambiguous than for hydrophobicity.

On the one hand, forest fires tend to produce greater water repellency in most forest soils, and this tends to reduce slaking, thereby increasing aggregate stability (MATAIX-SOLERA and DOERR, 2004).

On the other hand, aggregate stability is positively correlated with organic matter content, so the destruction of the organic compounds present in the soil tends to diminish aggregate stability (CERDÀ et al., 1995; GIOVANNINI et al., 2001).

The net effect probably depends on initial organic matter content, soil texture, fire intensity, and the size range of aggregates being considered.

(MATAIX-SOLERA and DOERR 2004) showed that finer sieve fractions (< 0.25 mm) are more hydrophobic than larger ones (> 0.25 mm), and that both hydrophobicity and aggregate stability are correlated with soil organic matter for different sieve fractions, but this question requires further research.

Soils on south-facing slopes tend to have both lower aggregate stability values and greater erosion rates than soils on north-facing slopes (CERDÀ et al., 1995; MARQUÉS and MORA, 1998; ANDREU et al., 2001).

Several reasons can explain this: south-facing slopes in the Mediterranean climate tend to have greater potential evapo-transpiration rates leading to sparser vegetation cover and therefore thinner soils and lower soil organic matter contents.

In addition, the drier conditions on south-facing slopes may make them more susceptible to forest fires with more frequent burns and exposure to soil erosion processes; finally, post-fire recovery on south-facing slopes is slower than on north-facing slopes (CERDÀ et al., 1995; PAUSAS and VALLEJO, 1999), so erosion processes continue for longer.

12.3 SITE DESCRIPTION

The study area is located in south-east France in the Massif des Maures near St Tropez; the approximate centre of the burnt area is 43°16’ N, 6° 28’ E.

The Giscle catchment has a surface area of about 234 km2 and is composed of two major topographical features.

The lower portion of the catchment (roughly 25% of the total area) is an alluvial plain occupied by vineyards and urban development.

The upper portion (about 75%) of the catchment has forested hilly terrain dominated by a mixture of Mediterranean oak and pine species.

A major forest fire which occurred at the end of August, 2003, burned more than 2000 ha located at the head of the Giscle river and a few of its tributaries.

Local authorities were therefore concerned about the impact of increased runoff on the urbanised areas downstream as well as the possible increase in sediment load entering the port.

Annual precipitation is about 950 mm with rainy seasons in the autumn and spring.

The hilly forested zone is underlain by metamorphic gneiss and schists, and the soils are classified as Rankers in the FAO classification.

Soil textures for the < 2 mm fraction are typically about 75% sand, 10% silt, and 15% clay. Mean Weight Diameter (MWD), measured using the method of Le Bissonnais 1996), for soils in the valley bottom is 3.28 mm (std. dev. = 0.07), classifying these soils as highly erodible.

EUFIRELAB

D-08-06.doc 116

12.4 METHODS

12.4.1 Mapping soil erosion risk

Since undisturbed forests typically have very low soil erosion rates, most soil erosion models were developed for agricultural contexts and were not suitable for the burned forest site.

Measuring soil erosion and elaborating a regression model were also out of the question due to the lack of time.

After an initial survey of the digital data available, the dominant factors affecting soil erosion (described below) were mapped and combined to create an erosion map.

The multiplicative method is described below after a presentation of the factors.

In the field, each layer was mapped on a 1:25,000 topographic map; some areas of the burned upper catchment were inaccessible except via footpaths, much of the burned catchment was accessible only via unmade roads (fire prevention access roads and private lanes), and only the lower portion of the catchment could be accessed by car.

Therefore, cycling throughout the catchment and combining direct on-site observations with panoramic views from viewpoints using binoculars for remote sites, carry out most of the mapping.

Roughly 2,000 ha had to be mapped in a few days, so each factor was attributed a number of qualitative classes that could be estimated quickly without time-consuming measurements.

The methods used are outlined below after a description of the soil erosion factors.

12.4.2 Soil erosion factors

Slope: the importance of slope inclination for soil erosion is well known, and a Digital Elevation Model (DEM) of the catchment with a 50 m grid was already available, but the resolution quickly revealed itself to be insufficient, so a 25 m DEM was ordered.

Pre-fire vegetation: pre-fire density was estimated from the standing charred trunks, and three categories were included – bare, low density, and medium density.

Rock outcrops were devoid of any vegetation and the distinction between low and medium density was based mainly upon the ease of walking through the charred forest.

In addition to the mulch effect provided by fallen pine needles described above (SHAKESBY et al., 1993; SHAKESBY et al., 1994), standing dead vegetation may favour infiltration (near stems and in burned roots) and slow runoff velocity, so the net expected effect is lower erosion where vegetation stands are more dense.

Fire severity: this is a major factor influencing runoff and soil erosion rates (ROBICHAUD and WALDROP, 1994, cited in LETEY, 2001; RUBIO et al., 1997; PROSSER and WILLIAMS, 1998).

Two severity categories were included: variable and intense.

Variable severity sites were designated as areas where:

- either the fire had passed quickly and charred the trunks and underbrush but left the crown more or less intact, or

- where the intense combustion of the standing vegetation was confined to a small area (dozens of square meters) In intense severity zones, the fire had consumed the

underbrush and crown entirely over an extensive area.

Soil erodibility: soil depth and stoniness distribution within the catchment are controlled by slope inclination, position within the catena, and aspect.

Thin stony soils were generally found on steep slopes, convexities, and south-facing slopes, while deeper less stony soils were found on gentler slopes, concavities, and north-facing slopes.

Three classes were defined: - thin and stony, - of intermediate depth and stoniness, and - deep with few stones.

A soil corer was used along a dozen slope transects and thin soils were designated as those that had depths of < 20 cm.

A soil was considered deep if it was > 50 cm in depth.

Ascribing coefficients and combining the soil erosion factors.

Mapping the soil erosion risk was based on the multiplication of coefficients (equ. 1) ascribed to the soil erosion factors summarized in Table 1.

Erosion Index = Slope x Vegetation density x Fire severity x Soil erodibility (1)

For each factor, a raster layer with a 25 m cell size was created using geographic information software.

Each cell within the layer was assigned a coefficient (Table 1) and the layers multiplied (equ. 1) to provide a theoretical range of values of from 1 to 135.

In the absence of a suitable, functional, and readily available soil erosion model, values for the coefficients were attributed arbitrarily according to personal experience and field observations.

Estimating coefficients for slope, vegetation and fire severity was relatively straightforward.

The soil erodibility coefficients were more problematic: shallow stony soils generate more runoff but surface material is less mobile due to the high stoniness, and in many cases, the bare bedrock.

Deep soils with fewer stones may have higher infiltration rates, but they are more erodible, particularly with respect to the weak aggregate stability and possible presence of hydrophobicity.

Therefore, we decided to increase the value of the coefficient with decreasing stone content, as can be seen in Table 1.

EUFIRELAB

D-08-06.doc 117

12.4.3 Partial validation of the soil erosion risk map

Once the soil erosion risk map was completed, the catchment was divided into sub-catchments and the percentage area in the high and very high risk categories (Table 2) was calculated for each sub-catchment.

On 2 December, 2003, a rainfall event generated significant runoff in the catchment: peak rainfall intensity values were not available but roughly 35 mm fell in the morning, and about 35 mm fell in the preceding five days (measured at about five km from the fire site.

Runoff samples from most of the major streams were obtained manually by plunging a 0.5 L bottle into the stream during the storm.

Replicate suspended sediment samples from the main channel and selected tributaries were collected within a period of about 10 minutes so rainfall/runoff conditions were considered comparable.

The amount of suspended sediment transported through the channel depends on both sediment concentration and discharge, but the timing imposed by sampling simultaneously and the difficult field conditions made it impossible to measure discharge accurately.

Therefore, we only measured suspended sediment concentrations.

An order of magnitude based on visual estimates would give discharges of about 10 m3 s-1 for the main channel and about 20-30% of that value for the tributaries.

It was therefore possible to compare the stream suspended sediment load to the percentage of the catchment in the high and very high-risk categories.

For comparison, we sampled a stream draining a vineyard sub-catchment and an unburned forest area.

12.5 RESULTS

12.5.1 Distribution of the soil erosion factors

Figures 1 to 4 show the distribution of the four factors of soil erosion in the catchment.

More than 30% of the slopes in the study area have an inclination greater than 20% (Figure 1), showing the natural vulnerability of the zone to erosion processes.

In addition, many of the steeper slopes were those most affected by the fire (Figure 2).

Typically, pre-fire vegetation densities (Figure 3) were low on convexities or south-facing steep slopes with shallow stony soils (Figure 4).

Conversely, deeper soils with a denser pre-fire vegetation cover tend to be concentrated in the southeast corner of the map where slopes are gentle.

12.5.2 The soil erosion risk map

Values resulting from the multiplication of the factors ranged from 1 to 90, and were subdivided into soil erosion classes according to breaks between peaks in the histogram distribution.

Table 2 presents the values for each class, and Figure 5 shows the results.

Each class tends to correspond to a typical environment (Table 2.

As stated above, the erosion model used is extremely simple and the subjective definition of the coefficients is open to some debate.

But the objective was to provide a quick spatial representation of general trends and not a quantitative estimate of soil loss.

So, it is unlikely that a more time-consuming modelling approach would have significantly altered the spatial distribution of the erosion classes identified.

The authors had been working in the catchment for three years prior to the fire, and soil erosion in the vineyards had been mapped using both the Revised Universal Soil Loss Equation (RUSLE) and a regression equation based on measurements carried out in a reference area.

Neither of these models, designed for agricultural contexts, could be applied to burned forest sites, and the data requirements for more complex deterministic models were unrealistic within the time frame allowed for the post-fire management proposals.

EUFIRELAB

D-08-06.doc 118

12.5.3 Partial model validation

Significant runoff occurred during the 2 December, 2003, storm (Figs 6, 7, 8. Stream discharge samples (2 replicates) were collected for three of the four main sub-catchments.

Sampling locations are shown in Figure 5: Tourré - measured just West of the “Pont de Bois” before

entering the main Giscle channel - the subcatchment surrounding the “Val de Gilly”

sampled just before entering the Giscle channel - the main channel sampled at the “Pont Gué”

(sampled upstream of other tributaries. The Périer in the southern part of the catchment could not be sampled in similar rainfall/runoff conditions, so it was not analysed. In addition, samples for the burned north-facing slope

- draining into the Giscle main channel just South of the “Pont Gué”; samples of runoff from a vineyard and an unburned forest were collected for comparison purposes.

For each of the three sub-catchments sampled and the burned north-facing slope, stream suspended sediment concentration was plotted against the sum of the percentage area in the high and very high-risk erosion classes (Figure 9).

The numbers in Figure 9 correspond to sampling locations described in Figure 5. Sediment concentration values represent the means of 2 samples.

Differences in concentration between samples were never greater than 10% except for the Gilly, for which it was 31%.

Although there are only four points, there is a general trend for the sediment concentration to increase with increasing area in the High and Very High erosion classes.

The exception is the Giscle (sample 3) which is located immediately downstream of a long straight section with natural sediment deposition.

Therefore, the low suspended sediment concentration of the Giscle was attributed to natural deposition in the stream bed upstream of the sampling area.

Sediment concentrations for the vineyard (5) and unburned forest were 1.64 g l-1 and 0 g l-1, respectively.

12.6 DISCUSSION

Identifying critical sites for post-fire erosion control methods depends on an accurate spatial estimation of high-risk zones.

Several factors contribute to make this task somewhat more difficult than in a typical agricultural setting.

These include the following: - The short intervention time between the forest fire

and the installation of post-fire erosion control methods. Identifying critical areas must be done within a period of weeks after the fire for the methods to be efficient during the first winter rains.

- The lack of spatial data on soil properties that contribute to enhance or reduce the erosion risk (soil depth, stoniness, hydrophobicity, aggregate stability…. Data for soil properties in agricultural areas tend to be more abundant than for forests since soil productivity depends on these properties. Data for forested areas are scarce.

- In the Mediterranean region, forested slopes tend to be particularly steep since flatter areas are occupied by agricultural activities, including vineyards, fruit and olive groves, and cereal crop. After a major forest fire, slope angle is perhaps the single most important factor determining erosion rates. DEM spatial resolutions should be in the order of 25 m to accurately represent the terrain.

- Pre-fire vegetation density and fire intensity are spatially variable and along with slope angle they play a major role in determining soil erosion risk.

The case study presented above describes an operational strategy to map erosion risk at the scale of a large catchment.

Work is currently underway to investigate two future directions of research.

The first is to elaborate a pre-fire strategy that would shorten the post-fire reaction time.

To what extent can planners prepare for the consequences of a forest fire before it occurs in order to respond more quickly to the potential erosion risk?

The second is to determine how the spatial scale of the data and model affect output.

Soil erosion risk maps have been produced to identify potential high risk areas at the European scale for essentially arable land.

The question now asked is “Can a similar approach be used for post-fire soil erosion risk?”.

EUFIRELAB

D-08-06.doc 119

12.7 CONCLUSIONS

Major forest fires in the Mediterranean environment occur regularly and might become more frequent in the future as the climate changes.

A method for choosing appropriate sites for the erosion control measures is based firstly on an assessment of the spatial distribution of erosion risk.

It should be noted that although the soil erosion risk map produced in the first stage of the method is useful in identifying high risk areas, it does not represent in itself a sufficient tool for developing a soil conservation strategy.

In the future, soil erosion modelling should be accompanied by the development of soil conservation models where high risk sites are evaluated for suitability to different erosion control methods.

Increasing the efficiency of the approach and the effects of varying spatial scale are two further research directions that are currently being pursued.

EUFIRELAB

D-08-06.doc 120

12.8 FIGURES

Figures 1 and 2: Slope classes (left) and Fire Severity (right)

Figures 3 and 4: Estimated pre-fire vegetation density (left) and distribution of soil characteristics (right)

Figure 5: Soil erosion risk map showing the erosion risk categories.

EUFIRELAB

D-08-06.doc 121

Figure 6: Runoff observed in the burnt forest during the 2 December, 2003, storm

Figures 7 and 8: Runoff flowing over a low bridge (Pont Gué) during the 2 December, 2003, storm (left)

and the same bridge (Pont Gué) at normal discharge (right)

20 40 60 80Area (%) in High & Very High Erosion

0

1

2

3

4

5

Sedi

men

t Con

cen t

ratio

n (g

L-1

)

1

2

3

4

4 GULLEY3 GISCLE2 GILLY1 TOURRE

Figure 9: Relationship between sediment concentrations in storm runoff and proportion of area (%) defined as high

or very high erosion risk (numbers refer to site locations described in Fig. 5)

EUFIRELAB

D-08-06.doc 122

12.9 TABLES

Table 1: Coefficients ascribed to the four soil erosion factors used to estimate erosion risk: slope, pre-fire vegetation density, fire severity, and soil erodibility

Slope (%) Pre-Fire Vegetation Density

Fire Severity Soil Erodibility

0-5 1 Medium 1 Variable 1 Shallow soil with high stone content 1

5-10 2 Low 2 Intense 3 Intermediate depth and stoniness 2

10-20 3 Bare 3 Deep soil with few stones 3

20-30 4

>30 5

Table 2: A brief description of the typical environments observed in the catchment for a range of erosion risk values

Class Erosion Value

Area km2 (%) Summary Description

Low ≤ 5 3.3 (16) Valley bottoms with gentle slopes, low severity fire, dense pre-fire vegetation and deep soils. Included are also some flat topslope sections where soils are stonier and shallower.

Intermediate 6 - 15 8.3 (40)

Convexities and concavities upslope and downslope, respectively, of major linear slope sections. These areas have gentler slopes than the linear sections and are frequently found along the limits of the pre-fire vegetation and fire serverity classes.

High 16 - 26 6.5 (31)

Steep slopes with bare or low pre-fire vegetation densities in the severely burned area dominate. One significant exception is the north-facing slope located south of the main Giscle channel. In this area, pre-fire vegetation density and soil depth are greater than elsewhere in the high erosion risk class, but the steep slopes make it vulnerable to erosion.

Very high ≥ 27 2.7 (13) Severe fire conditions in low density pre-fire vegetation (with patches of bare slope) on very steep slopes.

d-08-06[1]

Documents