international journal of remote sensing - tu wien · 2007-12-17 · landsat 7 etm+ remote sensing...
TRANSCRIPT
This article was downloaded by:[Swets Content Distribution]On: 16 November 2007Access Details: [subscription number 768307933]Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK
International Journal of RemoteSensingPublication details, including instructions for authors and subscription information:http://www.informaworld.com/smpp/title~content=t713722504
Detecting unknown coal fires: synergy of automatedcoal fire risk area delineation and improved thermalanomaly extractionC. Kuenzer a; J. Zhang a; J. Li b; S. Voigt c; H. Mehl c; W. Wagner aa Claudia Kuenzer, Institute of Photogrammetry and Remote Sensing (IPF), ViennaUniversity of Technology, A-1040 Vienna, Austriab Beijing Normal University, BNU, Beijing 100856, Chinac German Remote Sensing Data Center (DFD) of the German Aerospace Center(DLR), D-82234 Wessling, Germany
Online Publication Date: 01 January 2007To cite this Article: Kuenzer, C., Zhang, J., Li, J., Voigt, S., Mehl, H. and Wagner, W. (2007) 'Detecting unknown coalfires: synergy of automated coal fire risk area delineation and improved thermal anomaly extraction', International Journalof Remote Sensing, 28:20, 4561 - 4585To link to this article: DOI: 10.1080/01431160701250432URL: http://dx.doi.org/10.1080/01431160701250432
PLEASE SCROLL DOWN FOR ARTICLE
Full terms and conditions of use: http://www.informaworld.com/terms-and-conditions-of-access.pdf
This article maybe used for research, teaching and private study purposes. Any substantial or systematic reproduction,re-distribution, re-selling, loan or sub-licensing, systematic supply or distribution in any form to anyone is expresslyforbidden.
The publisher does not give any warranty express or implied or make any representation that the contents will becomplete or accurate or up to date. The accuracy of any instructions, formulae and drug doses should beindependently verified with primary sources. The publisher shall not be liable for any loss, actions, claims, proceedings,demand or costs or damages whatsoever or howsoever caused arising directly or indirectly in connection with orarising out of the use of this material.
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7
Detecting unknown coal fires: synergy of automated coal fire risk areadelineation and improved thermal anomaly extraction
C. KUENZER*{, J. ZHANG{, J. LI{, S. VOIGT§, H. MEHL§ and W. WAGNER{
{Claudia Kuenzer, Institute of Photogrammetry and Remote Sensing (IPF), Vienna
University of Technology, Gusshausstr. 27-29, A-1040 Vienna, Austria
{Beijing Normal University, BNU, Beijing 100856, China
§German Remote Sensing Data Center (DFD) of the German Aerospace Center (DLR),
Oberpfaffenhofen, D-82234 Wessling, Germany
(Received 18 August 2006; in final form 21 December 2006 )
This paper presents two complementing algorithms for remote sensing based coal
fire research and the results derived thereof. Both are applicable on Landsat,
ASTER and MODIS data. The first algorithm automatically delineates coal fire
risk areas from multispectral satellite data. The second automatically extracts
local coal fire related thermal anomalies from thermal data. The presented
methods aim at the automated, unbiased retrieval of coal fire related
information. The delineation of coal fire risk areas is based on land cover
extraction through a knowledge based spectral test sequence. This sequence has
been proven to extract coal fire risk areas not only in time series of the
investigated study areas in China, but also in transfer regions of India and
Australia. The algorithm for the extraction of thermal anomalies is based on a
moving window approach analysing sub-window histograms. It allows the
extraction of thermally anomalous pixels with regard to their surrounding
background and therefore supports the extraction of very subtle, local thermal
anomalies of different temperature. It thus shows clear advantages to anomaly
extraction via simple thresholding techniques. Since the thermal algorithm also
does extract thermal anomalies, which are not related to coal fires, the derived
risk areas can help to eliminate false alarms. Overall, 50% of anomalies derived
from night-time data can be rejected, while even 80% of all anomalies extracted
from daytime data are likely to be false alarms. However, detection rates are very
good. Over 80% of existing coal fires in our first study area were extracted
correctly and all fires (100%) in study area two were extracted from Landsat data.
In MODIS data extraction depends on coal fire types and reaches 80% of all fires
in our study area with hot coal fires of large spatial extent, while in another
region with smaller and ‘colder’ coal fires only the hottest ones (below 20%) can
be extracted correctly. The success of the synergetic application of the two
methods has been proven through our detection of so far unknown coal fires in
Landsat 7 ETM + remote sensing data. This is the first time in coal fire research
that unknown coal fires were detected in satellite remote sensing data exclusively
and were validated later subsequently during in situ field checks.
*Corresponding author. Email: [email protected]
International Journal of Remote Sensing
Vol. 28, No. 20, 20 October 2007, 4561–4585
International Journal of Remote SensingISSN 0143-1161 print/ISSN 1366-5901 online # 2007 Taylor & Francis
http://www.tandf.co.uk/journalsDOI: 10.1080/01431160701250432
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7 1. Introduction
Uncontrolled coal fires in coal seams as well as in coal waste and storage piles occur
worldwide and pose a great threat to the environment. Countries facing the problem
of coal fires include China, Russia, South Africa, the USA, Australia and Croatia,
amongst others. The fires lead to the emission of green house relevant and toxic
gasses, the deterioration of vegetation, land subsidence due to volume loss
underground, and to the loss of the valuable resource coal (Kuenzer et al. 2007).
Coal fires have been well studied with remote sensing methods based on multi-
sensor data. Documented remote sensing based coal fire research started around
1963, when a company producing airborne thermal cameras tested their equipment
above burning coal waste piles in Scranton, Pennsylvania. The thermal imagery
clearly showed the burning areas within the coal waste piles as light grey values
(Slavecki 1964). Further aerial surveys with thermal cameras in the 1960s and 1970s
were conducted by Moxham and Greene (1967), Greene et al. (1969), Knuth et al.
(1968), Fisher and Knuth (1968), Rabchevsky (1972) and Ellyett and Fleming
(1974).
From the 1980s, multi-spectral scanner data and thermal infrared scanner data
were used for the investigation of coal fires in the Chinese regions of Taiyuan Xishan
in Shanxi province (Guan 1989, Li 1985). Coal fire research based on aerial
multispectral scanner data continued in India, where Bhattacharya et al. (1991,
1994), Mukherjee et al. (1991) and Prakash et al. (1995) investigated the Jharia coal
field, generating coal fire maps derived from aerial thermal data based on ground
truth knowledge. Details on the capability of coal fire detection from airborne
thermal scanner data are presented by Zhang, X. et al. (2004). The Jharia region was
studied again based on satellite sensor data by Mansor et al. (1994), using thermal
and mid-infrared NOAA-AVHRR and thermal Landsat-5 TM data. Saraf et al.
(1995) as well as Prakash et al. (1997), Prakash and Gupta (1998) and Gupta and
Prakash (1999) applied Landsat-5 TM thermal band data with 120 m spatial
resolution in the same area for thermal anomaly extraction based on thresholding
techniques. Landsat-5 TM was often sufficient for detecting large surface fires and
shallow subsurface fires, but too coarse to detect very deep or small coal fires
(Zhang, X. 1998).
Landsat 7 ETM + based studies demonstrated slight advances compared to
Landsat TM due to the 60 m thermal band. Coal fire outlines can be mapped more
precisely than in TM data (Kuenzer, 2005, Zhang, 2004, Tetzlaff, 2004, Yang et al.
2005). Additionally, Kuenzer et al. (2004) presented regional studies analysing
multitemporal land cover change in the study areas, dating back to 1989.
Ganghopadhyay (2003) used ETM + data for coal fire mapping in China applying
local thresholds and Gao et al. (2006) compared pixel based and object oriented
classification methods in the coal mining environment of Wuda.
Generally, satellite data with a coarser spatial resolution than that of Landsat-TM
were rarely used for coal fire research. Though Mansor et al. (1994) could spot very
hot surface coal fires in Jharia, Zhang, X. (1998) showed that the spatial resolution
of NOAA-AVHRR data (1 km) is usually too low to detect underground coal fires
in Northwest China (Zhang, J. et al. 2004).
All the above mentioned remote sensing research focused on interactive analyses,
where coal fire thermal anomalies were extracted using density slicing (threshold-
ing). Spatially transferable or automated methods were rarely presented. Coal fire
research based on multispectral (non-thermal) remote sensing information has only
4562 C. Kuenzer et al.
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7 been for local descriptions of coal fire environments through manual landcover
classifications (Van Genderen and Guan 1997, Prakash and Gupta 1998, Prakash
et al. 2001, Gao et al. 2006, Kuenzer et al. 2004).
The goal of the present approach is thus to automatically extract multispectral
and thermal coal fire related information, and employ this information in a
synergetic way.
2. Study areas, data and pre-processing
The first study area, Wuda, is located in Inner Mongolia Autonomous Region,
China. The city borders a structural syncline that hosts a large coal-fire area that has
been investigated over the past several years (Li et al. 2005, Voigt et al. 2004, Zhang
2004, Tetzlaff 2004, Kuenzer 2005). It is located on the western side of the Yellow
River, north of the Helan Shan mountain range. The first dunes of the Badain Jaran
Desert are located 10 km to the west of Wuda. Wuda city centre is located at
approximately 39.51u North and 106.60u East. Elevation in this region varies from
1010 m to 1980 m above sea level.
The second study area, Ruqigou, is located in the northern part of the Helan Shan
mountain range in Ningxia Hui Autonomous Region. The Helan Mountains strike
from southwest to northeast for 200 km along the north-western border of Ningxia
and Inner Mongolia. Their average elevation is 2000 m; the highest peak, Helan
Mountain, reaches 3557 m. The mining area is centred at approximately 39.07uNorth and 106.12u East. Elevations in this region vary from 1400 m to 2640 m above
sea level. The location of both areas is shown in figure 1.
Coal has been mined in both study areas since the late 1950s. The first coal fires in
the Wuda area were discovered in 1961. Today 20 major surface and subsurface fires
are consuming coal in the structural syncline that is located west of the city. The fires
cover more than 4 km2, and the annual coal loss is estimated to exceed 200,000 tons
(Kuenzer 2005). The mines in Ruqigou have been affected by more than 20 coal
fires, most of which are currently being extinguished. The total area affected by coal
fires in the Ruqigou region is assumed to nowadays cover less than 2 km2 (Chen
1997; Kuenzer 2005). Field campaigns in the years 2002–2005 enabled the frequent
observation of coal fire dynamics through annual or bi-annual in situ mappings
(Kuenzer et al. 2005) as well as the collection of thermal and land cover ground
data.
Figure 1. Location of the study areas Wuda and Ruqigou in China.
Detecting unknown coal fires 4563
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7 Both study areas are contained within the Landsat-7 ETM + scene of path 133,
row 129. For this paper, a scene from 21 September 21 2002 was used for
multispectral and thermal analyses. Additionally, a corresponding night-time
Landsat-7 ETM + scene (path 226, row 211) of 29 September 29 2002 was used
for more precise thermal anomaly extraction, since solar effects are less accentuated
in night-time data. Furthermore, the only cloud free Landsat 5 TM scene from the
1980s for this area was analysed (path 129, row 33 on 20.09.1987).
Furthermore, three datasets were employed to demonstrate algorithm transfer.
These scenes cover the Jharia coal fields in Northeast India at approximately
23u429 N and 86u159 E (path/row: 140/044) and a mining area at the foothills of the
Flinders Range, Australia, at approximately 30u189 S and 138u369 E (path/row 098/
081). The inner-Chinese transfer scene is located over the city of Baotou in Inner
Mongolia (path 128, row 32, 29.08.2002).
All scenes from the Chinese study areas were geo-referenced based on precise
GPS-derived vector data from a field campaign in September 2002. The data were
also orthorectified to minimize terrain induced displacements. We used a DEM
retrieved from 25 m spatial resolution ERS-2 SAR data. Following, the data were
atmospherically corrected using the ATCOR-3 code (Richter 1998). This code also
normalizes topography induced illumination effects. Working with reflectance data
enables the inter-comparability of scenes from different times and areas. This is
especially mandatory for the automated delineation of coal fire risk areas, while the
thermal algorithm can run on pre-processed or raw data equally well (table 1).
3. Methodology
3.1 Automated delineation of coal fire risk areas
The focus of the present approach lies on multispectral data analyses to delineate
areas, where coal fires are likely to occur. Based on field observations and mappings
during fieldwork campaigns in 2002, 2003, 2004 and 2005 it was found that coal fire
related thermal anomalies are always located adjacent to coal exposed to the surface
(Kuenzer 2005). At risk for coal fires are abandoned and active mines, outcropping
coal seams, coal waste piles, coal storage piles, mining portals or coal washery
discard. Hence, the detection of the surface class ‘coal’ and knowledge of the
distribution of coal is crucial for the delineation of possible coal fire and coal fire
risk areas (Kuenzer 2005). Many coal fires occur underground and the thermal
anomalies may not be located directly ‘on top’ of a coal surface (e.g. a sandstone
layer, which is overlaying a burning coal seam). However, coal fires in China occur
Table 1. Data sets investigated for the presented research. The main test areas Wuda andRuqigou, Gulaben are both contained in data set path 129/033. The lower three scenes
represent transfer areas in China, Baotou, Australia and India respectively.
Path/RowAcquisition
dateLatitude
[UL corner]Longitude
[UL corner]Sun elev.
[u]Sun azimuth
[u]Day/night
129/033 20 Sep 1987 39.86uN 105.57uE + 47.00 + 146.26 day129/033 21 Sep 2002 39.87uN 105.58uE + 47.02 + 147.12 day226/211 29 Sep 2002 39.32uN 104.40uE / / night128/032 29 Aug 2002 41.28uN 107.45uE + 52.84 + 139.80 day098/081 15 Nov 2002 29.33uS 137.42uE + 59.53 + 074.15 day140/044 2 Nov 2001 24.05uN 087.81uE + 46.31 + 147.98 day
4564 C. Kuenzer et al.
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7 not deeper than approximately 150 m underground. Below this depth the fires either
do not receive enough oxygen to keep burning or groundwater is present.
Furthermore, very deep fires do not lead to detectable surface thermal expressions.
Under this assumption, a buffer zone around the detected coal surfaces can be
derived, which includes the area where coal fire thermal anomalies are likely to
occur. The size of this buffer zone depends on the dip angle of the coal seam. The
farthest horizontal distance of a possible underground coal fire (and its surface
thermal anomaly) relative to the location of the outcropping coal seam can be
calculated with Dpmax[m]5Bf[m]6tan cda. Here Dpmax is the maximal coal fire depth
in metres (constant), Bf is the delineation radius in metres and cda is the dip angle of
the seam (see figure 2).
Coal fires lead to the degradation and decay of vegetation. Subsurface fires are
not found underneath densely vegetated soil or bedrock (Kuenzer and Voigt 2003,
Gupta and Prakash 1998). Very hot coal fires lead to the pyrometamorphosis of the
surrounding bedrock. If the thermal intensity of a fire is high enough colour and
texture changes of the adjacent strata can occur (Zhang 1996). These two
phenomena can be detected in remote sensing data. They are also considered in
the knowledge based test sequence, which was employed for coal surface extraction
and is presented in the following. However, since they are not the key element for
the delineation of risk areas and since pyrometamoprhic areas (and influenced
pixels) are rare, this is not presented in detail here (see figure 3).
The sequence automatically extracts coal surfaces from multispectral data.
Following, the area in a certain radius is delineated to yield possible coal fire risk
areas. Within these risk areas densely vegetated areas can be excluded and pixels of
pyrometamorphic rock can be highlighted. The spectral test sequence is based on the
principle of decision trees, which are a simple and transferable method to classify
image data without user interaction once the tree has been designed (Cihlar et al.
1998, Friedl and Brodley 1996). The spectral tests and the individual thresholds were
defined based on extensive statistical analysis of spectral reflectance signatures from
Figure 2. Delineating the area, in which coal fire related thermal anomalies can be expected.The maximum depth of the fires is a rough estimate. In China, no fires located deeper than150 m, while still leading to thermal anomalies on the surface, were reported.
Detecting unknown coal fires 4565
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7
a larger number of Landsat scenes. This grants temporal and spatial transfer of the
sequence.
Figure 4 demonstrates the working steps of the algorithm. A complete ETM + or
Aster layer stack is used to generate several intermediate data products. These are
spectral mean value channels (SPM) and vegetation indices (here SAVI, the soil
adjusted vegetation index for arid regions, Purevdorj et al. 1998) followed by testing
for the three surfaces of interest as mentioned. Following, the test sequence for
Landsat 7 ETM + data is presented.
3.1.1 Coal extraction: Test 1: Synthetic spectral mean test SPM123457. From all
input bands except the thermal channel a synthetic band is calculated, containing
the spectral mean (SPM) of each pixel over all bands. This mean channel gives an
indication of the albedo of a surface over the whole covered wavelength region.
Surface classes like coal, clear water or shadow have low spectral mean values due to
their overall low reflectance (r [%]) in the visible (VIS), near infrared (NIR) and
mid-infrared (MIR) domain. These three are therefore also the surface classes easily
to be confused. Especially coal and shadow have near identical spectra and even
illumination correction cannot eliminate all shadowed areas within satellite data.
r %½ �SPM123457§1 and r %½ �SPM123457ƒ10 ð1Þ
3.1.2 Coal extraction: Test 2: Synthetic spectral mean test SPM123. Additionally, a
SPM was calculated over the visible bands 1, 2 and 3. For the synthetic mean
channel of bands 1, 2, and 3 a lower threshold was defined at an average reflectance
of 1% and an upper threshold was defined at 7% reflectance.
r %½ �SPM123§1 and r %½ �SPM123ƒ7 ð2Þ
Figure 3. Radiometrically corrected Landsat-7 ETM + spectra for different surfaces (here:65MIR, 75TIR to present channels in the correct wavelength range). The image spectra ofcoal (black diamonds) and shadow (black squares) run nearly parallel and can only bedifferentiated by temperature differences. Water (grey diamonds) shows a similar spectralbehaviour in the first four bands but differs in the far NIR and MIR. The separability of coal,vegetation, or geologic surfaces, like sandstone, limestone and desert sand (upper threesignatures) is good. The right part shows average reflectance and standard deviation of coal(left), shadow margin areas (second left), shadowed areas (second right) and water of theYellow River (right). Especially coal and shadow margin regions can be easily confused. Here,the thermal band can support separation of classes.
4566 C. Kuenzer et al.
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7
These first two thresholds help to delineate coal surfaces from other geologic
surfaces. Shadow, shadow margin areas and deep water cannot be separated yet.
3.1.3 Coal extraction: Test 3: Shadowed area tests. To eliminate conflict situations,
where shadow pixels are still mistakenly extracted as coal after the two SPM, the
following tests are applied:
Shadow file from ATCOR-3~0 ð3Þ
r %½ �band 1~0 ð4Þ
r %½ �band 2~0 ð5Þ
r %½ �band 3~0 ð6Þ
The first test includes the shadow-file, which is created during illumination
correction in ATCOR-3 (Richter 1998). This test is similar to the retrieval and
exclusion of shadowed areas out of a DEM-derived illumination file directly. Fully
shadowed regions (shadow file50) are defined as areas, where the pixel consists
purely of shade and the geometry between the pixel (relief), Sun and satellite sensor
does not allow for a return signal in these wavelengths. In contrast, other shadowed
pixels may only contain part of a shadow (shadow margin) and, therefore, allow for
Figure 4. Summarized workflow of the knowledge based test sequence. Marked in grey: testsequence for coal surface extraction. From the original input data set (upper left corner) thesynthetic bands (spectral means and vegetation index) are calculated. The data set is stacked.Then the test sequence for coal surface extraction (grey) is employed. If available, DEMderived input data supports this procedure (left incoming branch).
Detecting unknown coal fires 4567
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7 a weakened spectral signal. Such shadowed areas, which are not excluded in the
shadow file, can fulfil all or some of the equations (4)–(6). They can therefore be
excluded and will not be interpreted as coal surfaces. Only pixels not fulfilling any of
the equations will continue in the test sequence.
3.1.4 Coal extraction: Test 4: NIR test. This test excludes pixels whose reflectance
is too high in the NIR band 4 to belong to the class coal. Pixels which pass this test
continue in the testing sequence, pixels which do not pass the test are excluded from
further analysis.
r %½ �band 4ƒ10 ð7Þ
3.1.5 Coal extraction: Test 5: Test for water surface exclusion. This test supports
the exclusion of water surfaces, which might still be left within the remaining pixels,
which passed tests 1 and 2 and 4 and were not excluded by test 3. Water surfaces
often have low overall albedo and, therefore, may pass tests 1 and 2. Their
differentiation criterion to coal is that they usually exhibit lower reflectances in band
5 and band 7. Here, the spectrum of water has decreasing reflectances from blue to
the MIR.
r %½ �band 5§5 ð8Þ
r %½ �band 7§5 ð9Þ
r %½ �bands 4, 5, 7w0 ð10Þ
Pixels which do pass this test continue in the test sequence, pixels which do not fulfil
the criteria of 8 and 9 are assumed to be water surfaces. If 10 is fulfilled these are
either shadow surfaces (which were already excluded by test 3) or water surfaces ofextremely low reflectance.
3.1.6 Coal extraction: Test 6: Temperature test. The inclusion of the thermal band
can increase the separability of classes. The thermal band (calibrated and
radiometrically corrected to land surface temperature, LST) is included in the testsequence. This test excludes pixels which passed test 1 and test 2, did not pass test 3
and passed tests 4 and 5, but whose temperatures are not suitable for coal surfaces.
A problem with this test is that the temperature behaviour of a surface changes over
the year. Daily and annual amplitudes of a surface are not only a function of the
surface’s thermal inertia but also of longitude and latitude, elevation, microclimatic
location and strongly depend on the season’s climate, shading, wind, snow cover,
etc. It is not possible to define a fixed absolute temperature threshold applicable for
several regions and multi-date imagery. Therefore, a relative temperature test wasdefined. This relative approach is transferable from scene to scene. It might not yield
as exact results as an absolute threshold but allows for multi-scene processing.
r 0C½ �wGlobal mean band 6 0C½ � ð11Þ
Here r is emission of thermal radiation, which should not be confused with
emissivity. Condition 11 tests if the pixel under investigation has a temperatureexceeding the global mean of the thermal band. This is based on the finding that
shaded pixels (cooler) are always located left (at lower temperatures) of the global
4568 C. Kuenzer et al.
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7 mean of band 6, whereas coal pixels are located right (at warmer temperatures) of
the global temperature mean.
Pixels which pass tests 1, 2, 4, 5 and 6 and which do not pass test 3 are defined as
coal pixels. The final result of the above knowledge based coal extraction
methodology is majority filtered with a 363 matrix. This statistical filter excludes
single scattered pixels. Following this filter application, a buffer region is generated
around the extracted coal surfaces. The radius of the buffer depends on the prevalent
dip angle in the area. If no information on the general dip of the strata is available, the
algorithm runs with a default buffer size of 500 or 1000 m, depending on analyst
choice. With such a radius prevalent dip angles of 16u and more and 8.5u and more are
covered. However, this buffer size can be adjusted depending on test area.
3.2 Automated thermal anomaly extraction
Many research disciplines that work with remote sensing data define thresholds
empirically, meeting the needs to extract the desired information. One example is the
automated extraction of thermal anomalies resulting from forest fires. Worldwide,
many algorithms exist to extract forest fire related thermal anomalies from remote
sensing satellite data such as MODIS, NOAA-AVHRR or METEOSAT. They are
usually based on temperature thresholds, above which a pixel will be declared as fire.
It needs to be underlined that the task to detect a coal fire related thermal
anomaly against a normal temperature background is much more complex than for
example detecting a forest fire signal. Coal fire surface expressions are very subtle.
Even if a 30 cm long crack in an area of underground coal fires emits 400uC hot
gasses, the overall temperature of the 60660 m (or coarser) thermal pixel might only
be raised a few degrees Celsius against the background. Thus, coal fire anomalies are
extremely weak anomalies, which can by no means be compared with thermal
applications like forest fire detection, lava flow detection or the spotting of large
industrial heat islands. Coal fire thermal anomaly detection, and here especially
subsurface coal fire detection, is considered a very difficult task. In recent decades,
coal fires were thus mainly delineated by manual thresholding techniques, where the
fire outline was already known (e.g. through in situ field mapping or from coal fire
maps provided by mines). The threshold was then adjusted, so that the portion of
the thermally anomalous area derived in this process fit the known fire outline as
well as possible. Such interactive analyses can lead to very detailed coal fire maps,
but only if other such detailed coal fire maps already exist. Therefore, it is often
questionable, which additional or new information is gained from the Earth
observation data. Furthermore, one single threshold cannot be applicable to all the
coal fires within one satellite image and hence the image needs to be split into several
subsets to extract each fire with its own optimum threshold. The disadvantage of
these interactive analyses is very time consuming processing. Results can hardly be
repeated by a second analyst (interpreter’s bias) and the additional information gain
is relatively low.
The algorithm presented here circumvents the above disadvantages. It facilitates
raw satellite data as input for a sub-image statistical analysis and is based on a
moving window concept. For an image of M columns and N rows a small window
with M1 columns and N1 rows is defined (M1,5M, N1,5N). The window moves
over the image with the step X, where (X,5M-M1) is in the column direction and
step Y (where Y,5N-N1) is in the row direction. When the moving window passes
the whole image, a pixel in the image will be sampled in (M16N1)/(X6Y) subsets.
Detecting unknown coal fires 4569
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7 This moving window of varying size (depending on sensor of input data, usually
between 11611 and 35635) investigates the histogram of the image subsets
concerning the occurrence of thermally anomalous pixels. Based on the assumption
that each histogram of such a sub-image is the combination of a background part
and a coal fire part, the thermally anomalous part is separated from the background
part.
To not employ empirical thresholds a clearly defined mathematical point has to
be chosen for separation. In this case the first local minimum after the main
maximum (LM in figure 5) could be defined as the criterion allowing the separation
Figure 5. A moving window filter of varying size (11611 to 35635) investigated thethermal data set. The histogram of each sub-window can contain no thermal anomaly at all—gradually being up to 100% filled with a thermal anomaly as presented above in synthesizedimage subsets. We assume that every sub-window histogram contains a part representing thebackground temperatures (B) and the thermally anomalous part (F). The first local minimumafter the main histogram maximum (LM) is defined as the relative threshold to separate thosetwo. This automated method will lead to a loss of thermal anomalies if e.g. the sub-window is100% filled by a thermal anomaly. It will furthermore indicate coal fire or thermal anomaliesrespectively if no thermal anomalous area is contained in the image. However, since everycentre pixel is investigated for over 1000 times and has to be declared as ‘thermallyanomalous’ in at least 70% of the tests, this error is kept as low as possible. The principle ofthe algorithm is shown in the lower left area of the figure, the lower right area shows clusteredthermal anomalies (pixel size 60 m) above subsurface coal fires in Wuda (left) and above coalindustrial plants (right).
4570 C. Kuenzer et al.
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7 of thermally anomalous pixels from background pixels (see figure 5). In a moving
window approach each centre pixel within this window matrix is sampled multiple
times (depending on window size, which varies). In the present case the window sizes
11611, 19619, 27627 and 35635 were applied. One pixel is thus investigated
121 + 361 + 729 + 122552436 times. Only if a pixel is regarded as thermally
anomalous in 70% of the cases (.1706 times), it is declared as a thermal anomaly.
In this way, strongly contrary to an overall threshold, the thermal anomalies
extracted represent local anomalies. This means that pixels of very different
temperature and within a different temperature background can be declared
thermally anomalous. This method thus considers the large variation in coal fire
anomaly temperatures and locations. Figure 6 shows histograms for three coal fire
areas—fire #8 in Wuda syncline, fire # 41 in Ruqigou and fire # 21 in Gulaben. All
three fires have different minimum and maximum pixel values as well as a differing
temperature range in thermal night-time data.
From all extracted pixels many anomalous pixels do not stem from coal fires; such
as for example a relatively warm small lake located within a cooler surrounding
background. To remove such false alarms the first output set of thermal anomalies
undergoes a statistical post-processing. The thermal anomaly’s direct neighbour-
hood is investigated for adjacent anomalies. If these exist the thermal anomalous
pixels are then clustered based on an eight-pixel neighbourhood. Each cluster’s
minimum-, maximum- and mean digital value, standard deviation and its spatial
coverage is written out into a file. These statistics are then compared with a
Figure 6. Histogram for thermal anomalous areas on a night-time ETM + image acquiredin 2002. Different thermal anomalies related to coal fires have different start and end DNvalues (respectively temperatures).
Detecting unknown coal fires 4571
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7 knowledge database of typical coal fire statistics generated during numerous field
campaigns. In China, no fire areas exceeding 1 km2 are known. If the algorithm
picks out an anomalous cluster above 300 60 m pixels (a little above 1 km2) it is
probably a strongly Sun illuminated slope. We choose 1 km2 as a threshold, since
coal fire related anomalies larger than 0.5km2 have never been reported in Landsat
or Aster data. To the authors no publication is known where single coal fires in situ
extend for over 1 km2. Furthermore, if the cluster’s temperature variance is too low
or nearly uniform it is not likely to be a coal fire cluster, where we would expect a
high variance. From five years of field survey we know that the coal fire we
investigated (and present in this paper) shows a high temperature variance. Thermal
clusters, whose variance is lower than that of the surrounding background in a
buffer zone of five pixels around the cluster, are rejected. With this careful post-
processing (sequence shown in figure 7), the resulting binary output file contains
thermal anomalies of small size, having a more reasonable chance of being coal fires.
Still, anomalies remain, which are not related to coal fires and which also cannot
be distinguished from a coal fire anomaly. Such anomalies are smaller Sun
illuminated surfaces, which pass the aerial threshold mentioned above (even
occurring in 10 pm night-time data), anomalies resulting from industry, the heating
of houses, burning of fields or garbage, or even forest or grassland fires. Here, the
previously described algorithm for coal fire risk area delineation (Kuenzer 2005)
comes into play. The synergy of both algorithms is presented in the results section.
The thermal algorithm works fully automatic for Landsat 5 TM, Landsat 7
ETM + , ASTER and MODIS data. The latter have been investigated by Hecker and
Kuenzer (accepted). They demonstrate its capability for coal fire detection in
MODIS data recently available as very valuable pre-dawn data from the platform
AQUA. Ratio images (band 20 versus band 32) were used for the detection of
surface fires.
4. Results
4.1 Results of automated coal fire risk area delineation
Figures 8 to 10 show the intermediate results and the temporary output image files
of the knowledge based test sequence for the automatic extraction of coal surfaces
(see figure 8). The result of the first spectral mean test is shown in figure 8b. It is
obvious that only very dark surfaces, such as coal, shadows and some very dark
water pixels remain. Especially the deeply incised valleys in the East of the subset as
well as smaller shadowed areas in the central limestone complex and the northern
foothills of the Helan Shan (lower left corner) can be seen. Such shadowed areas are
Figure 7. Sequence of the algorithm for automated thermal anomaly extraction. Movingwindow approach, anomaly clustering and statistical post-processing.
4572 C. Kuenzer et al.
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7
removed by test 3. Figure 9 impressively shows the automated extraction and
rejection of shadowed areas for complex terrain in the Ruqigou area. However, even
after this correction shadowed pixels (especially shadow margin areas) sometimes
still remain. These remaining scattered pixels can be seen in figure 10a. They can be
rejected through the temperature test 6. The remaining image (figure 10b) is a binary
file with coal surfaces displayed in white. This result undergoes a majority filtering
to exclude the few remaining single pixels, which are not representing coal surfaces.
We cannot exclude that some coal surfaces might consist of one pixel (which would
be eliminated by the filter)—with the filter we avoid the problem of retrieving a
spatially too large coal fire risk area. A buffering of the coal areas then leads to the
generation of so-called coal fire risk areas.
Figures 11–13 shows the result of coal extraction for the transfer areas in Baotou,
China, Jharia coal mining area, India, and for the Leigh Creek coal mining area in
Australia. Coal surfaces are generally shown in white. The output of the test
Figure 8 (a–b) Synthetic spectral mean channel over bands 1, 2, 3, 4, 5 and 7 (a); result afterthe test 1: pixels which passed the test are presented in white and include coal, shadow andwater pixels (b). Centre coordinates: 650705 E, 4373287 N, UTM, Z48 N.
Figure 9 (a–b) Visible band displayed in grey scale mode (a) and shadowed areas extracted bytest 3 supported by available illumination data derived from a DEM. Here presented for thestrongly shadowed area of Ruqigou. Centre coordinate: 607794 E, 4336506 N, UTM, Z48 N,Subset extent: 45 by 35 km.
Detecting unknown coal fires 4573
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7
sequence are binary image files with black areas (0) indicating background, while
white areas (1) are areas which passed the test sequence. Extraction accuracies were
assessed through the measure of producer accuracy based on ground data available
from different sources. Extraction accuracies are shown in table 2.
Figures 14 and 15 present the extracted coal surfaces in the regions of Wuda (10.1)
and Ruqigou (10.2) for the years 1987 and 2002, respectively. For the Wuda area it
is obvious that outcropping and exposed coal in the mining syncline (left part of the
image) has substantially increased. We also highlighted a coal waste pile (see
zoomed area) which has steadily grown over time. Furthermore, the dotted features
at the right side of the image have expanded. These are private small wild mines
along an outcropping seam running parallel to a valley road. See figure 8 for
location details. With a strong increase in exposed coal surfaces, coal fire risk has
also risen substantially. The same can be seen in figure 15. The coal mining area of
Ruqigou has expanded (lower left corner). Furthermore, in the two parallel valleys
Figure 10 (a–b) Pixels which passed tests 1, 2, 4 and 5 and were not excluded by test 3 (white)in the left image (a), result after the final relative temperature test 6 (b), centre coordinate:650705 E, 4373287 N, UTM, Z48 N. Subset extent: 45 by 35 km, as in 7.1.
(a) (b)
Figure 11. Original data (a) and coal pixels extracted for the scene of Baotou, northernChina (b). The subset displays only part of the full scene. Centre coordinate: 404361 E,4502452 N, UTM, Z49 N, WGS 84. Extent of subset: 30 km by 23 km.
4574 C. Kuenzer et al.
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7
(a) (b)
Figure 12. Original data (a) and coal pixels extracted for the Jharia coalfield area in north-eastern India (b). The subset displays only part of the full scene. Centre coordinate: 428060 E,2624047 N, UTM, Z45 N, WGS 84 Extent of subset: 35 km * 25 km. Available in colouronline.
(a) (b)
Figure 13. The Leigh Creek coal mining area in Australia (a), coal pixels extracted for thescene (b), this subset displays only part of the full scene. Centre coordinates of subset:6626009 S; 252463 E, UTM, Z54 S, WGS 84. Extent of subset: 60 km * 50 km. Available incolour online.
Table 2. Producer accuracy of the extracted coal (filtered) and dense vegetation pixels forWuda area, Ruqigou area and the transfer regions in China, India and Australia.
Date Prod. accuracy, coal (%) Prod.accuracy,densevegetation(%)
1987–09 Wuda not available not available2002–09 Wuda 92.10 (mapping) 95.04 (mapping)1987–09 Ruqigou not available not available2002–09 Ruqigou 92.10 (mapping) 95.04 (mapping)2002–08 Baotou 92.22 (landcover maps) 96.64 (landcover maps)2001–11 Jharia 87.51 (landcover maps) 99 (landcover maps)2002–09 Leigh Creek 98.03 (mine maps) not available
Detecting unknown coal fires 4575
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7
of Hulusitai and Shitanjing no mining activities could be spotted in the 1980s.
Fifteen years later exposed coal can clearly be spotted in the 2002 output image.
Figure 16 presents the panchromatic channel of the Landsat 7 ETM + daytime
scene from 21 September 2002. Coal fire risk areas (buffers around the extracted
coal surfaces) are superimposed on the image. Note that only a very small area
(,2%) of the whole dataset is considered as a coal fire risk area. Thermal anomalies
found outside these risk areas can be rejected as possible coal fire related thermal
anomalies.
4.2 Results of automated thermal anomaly extraction
Figure 17 shows the result of thermal anomaly extraction with the automated
detection algorithm based on Landsat 7 ETM + night-time data from 28.09.2002
superimposed on the panchromatic band of Landsat ETM + daytime data from
2002. It is obvious that not only coal fire related anomalies have been extracted. The
coal fire areas of Ruqigou and Wuda are represented well. Nevertheless, several
other anomalies outside the known coal fire areas exist. The number of such false
alarms can be minimized when utilizing the algorithm for the delineation of coal fire
risk areas as presented in the next section.
Figure 14. Exposed coal surfaces (white) in the Wuda area, 1987 and 2002. Available incolour online.
Figure 15. Exposed coal surfaces (white) in the Ruqigou area, 1987 and 2002. Available incolour online.
4576 C. Kuenzer et al.
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7
Figure 16. Coal fire risk areas superimposed on the panchromatic band of Landsat 7ETM + scene of 21 Sep 2002. Upper left corner: 39.87 N, 105.58 E, scene extent: 185 km *185 km.
Figure 17. Thermal clusters extracted with the automated detection algorithm based onLandsat 7 ETM + night-time data from 28 Sep 2002 superimposed on the panchromatic bandof Landsat ETM + daytime data from 21 Sep 2002. Detected anomalies: w) Wuda coal fires,r) Ruqigou coal fires, c) other confirmed coal fires, i) industrial installations. Extent of image:80 km by 90 km.
Detecting unknown coal fires 4577
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7 A similar result is shown for a MODIS pre-dawn dataset from 2005 to
demonstrate the multi-sensor capability of the thermal algorithm.
4.3 Combination of the two algorithms and synergetic effects
The powerful combination of the multispectral surface extraction algorithm for coal
fire risk area delineation with the thermal anomaly extraction algorithm could be
demonstrated, when new coal fires could be detected from remote sensing data
exclusively (see sections 4.3.1 and 4.3.2. below).
To quantify the synergetic effect of the coal fire (risk) area delineation algorithm
and the thermal anomaly detection algorithm, it was furthermore investigated how
many false alarm anomalies can be rejected through coal fire (risk) area delineation.
This is shown in table 3.
From the six scenes processed (five daytime, one night-time) we calculated howmany thermally anomalous clusters were extracted, how many anomalies were
rejected as false alarms through a clipping process with the delineated risk area (one
time with a 500 m buffer, one time with a 1000 m buffer), and how large the resulting
thermally anomalous area is. The thermal night-time data set yielded 243 or 211
remaining thermal anomalous clusters in the scene covering the two main study
areas. The area of these clusters covers 7.54 km2 and 6.77 km2 respectively. From
field mappings and local mining information it is known that coal fires in Wuda
cover an area of roughly 4 km2 and in Ruqigou of around 2 km2. Based on this resultand through visual checks and ground validation we know that approximately 90%
of the night-time thermally anomalous clusters within the delineated risk areas stem
from coal fires. Only a small number of these clusters stems from coal mining related
industry and other industries or settlements.
From the daytime scenes an unevenly higher number of thermal anomalous
clusters was extracted. This demonstrates the much higher need for the application
of the delineation algorithm for daytime data. The scene from the study areas in
1987 yielded 9817 anomalous clusters. In 2002 it was already 10900 clusters.
Illumination conditions were similar, since both scenes were acquired during thesame day of the same month. From weather station data we know that both days
were sunny and clear. Atmospheric differences will anyhow not influence single
Table 3. Coal fire risk area delineation supporting the exclusion of false alarms.
Anomalies in #pixels extracted bythermal algorithm
Anomalies in #pixels rejected
by coal fire riskarea algorithm
Anomalous areabefore coal firearea delineation
(km2)
Anomalous areaafter coal fire
area delineation(km2)
226/211 ETM +night 28 Sep 2002
451 243 (500m)211 (1000m)
14.49 6.77 (500m)7.54 (1000m)
129/033 ETM +day 21 Sep 2002
10,911 9449 (500m)8884 (1000m)
61.46 10.77 (500m)13.22 (1000m)
129/033 ETM +day 20 Sep 1987
9,817 9,078 (500m)8593 (1000m)
100.41 8.48 (500m)11.22 (1000m)
140/044 ETM +day 2 Nov 2001
6,067 4533 (500m) 42.84 14.92 (500m)
128/032 ETM +day 21 Aug 2002
15,653 15036 (500m) 83.90 3.90 (500m)
098/081 ETM +day 15 Nov 2002
1334 1300 (500 m) 7.25 0.14 (500 m)
4578 C. Kuenzer et al.
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7 pixels reflectance values so strongly that they would pass as a coal pixel (which has
nearly 0 reflectance except in the thermal band). The increase in anomalies thus
stems from the extension of the two presented coal fire areas, the ignition of new
coal fires in new areas and an increase in industry and settlement in the region.
Through risk area delineation, over 80% of the anomalous clusters were rejected as
false alarms However, even after the clipping with risk areas, the number of
remaining thermally anomalous clusters was still higher in daytime data. This results
from remaining solar effects and higher industrial and heat related activities during
the daytime.
However, it should also be mentioned that for Wuda not all coal fires were detected
by the thermal algorithm. Even in night-time data only 14 of the, at that time (2002)
existing 17, coal fires were detected. Three coal fires did simply show too weak thermal
signals to be extracted. These fires are usually fires which are very small in extent or
located very deep underground and thus do not yield a strong enough rise in
temperature of the overall 60 m pixel. It should be underlined though that even with
manual thresholding these fires can usually not be extracted. The same applies for the
extraction from MODIS data (see figure 18). In MODIS 1 km pixels only the hottest
coal fires will lead to sufficiently elevated signal and for Wuda only three of at that
time 20 (2005) coal fires were extracted. These are the hottest fires in the syncline. In
Ruqigou, where fires are larger and very hot, 80% of all coal fires were extracted even
in MODIS. In Landsat data of the Ruqigou study area all existing coal fires were
extracted correctly. To present detectability limits of coal fires depending on their size
and temperature would exceed the scope of this paper. Detailed calculations on this
topic can be found in Tetzlaff (2004). The work flow of combined risk area delineation
and thermal anomaly detection is prevented in figure 14.
Figure 18. Thermal clusters extracted with the automated detection algorithm based onMODIS ratioed pre-dawn data from 2 Feb 2005. Detected anomalies: w) Wuda coal fires, r)Ruqigou coal fires, c) other confirmed coal fires, i) industrial installations, y) Yellow River.Ground extent (field mapping) of Wuda and Ruqigou coal fires are indicated as polygons inwhite and dark grey, respectively. Background image: Landsat 7 ETM + from 21 Sep 2002,band 1. Extent of image: 80 km by 90 km.
Detecting unknown coal fires 4579
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7
4.4 Detection of unknown fires: field survey in September 2003
In September 2003, a field trip was undertaken to the remote area shown in figure 20.
The coal-fire fighting team of Wuda, which joined the survey, did not know about the
area visited beforehand, nor did they have any information on possible coal fires in
that region. The area is settled very sparsely and is only accessible via dirt roads.
Information from the local inhabitants revealed that the region once was a prospering
coal mining area. However, production had decreased drastically because of financial
problems. With the help of GPS, the satellite data thermal anomalies were located in
an area of very rugged terrain of former coal waste piles, abandoned mines, and
heavily disturbed coal outcrops. Five of the six suspicious thermal anomalies (as
extracted from thermal satellite sensor data and located within risk areas) were fires in
a coal seam, whereas the other anomalies resulted from fires in a coal waste pile.
Therefore, all six anomalies could be verified as coal fire anomalies. In this area, the
fire temperatures ranged from 170uC to 340uC.
4.5 Detection of unknown fires: Field survey in June 2004
In June 2004, another field trip was undertaken to the eastern side of the Yellow
River, approximately 25 km southeast of Wuda (see figure 21). Based on remote
sensing data exclusively, this area was expected to host at least one coal fire. The
thermal anomalies that were extracted and located within the delineated areas were
fires in underground coal seams. The local inhabitants knew about the fires; however,
they were unknown to the mining authorities of Wuda city. According to a local
worker, the fire had developed in private coal mines that were abandoned and sealed
improperly. The local people started to seal these mines with sand and loess just
recently. Coal-fire fighting is currently undertaken. A trench is dug in an attempt to
separate the burning portion of the coal seam from the unaffected part (see figure 21).
To date the fire is not under control and spreads along the dip and strike directions.
4. Conclusions
The presented algorithm for coal fire risk area delineation is based mainly on the
automated extraction of the surface class coal. The algorithm is spatially and
temporally transferable and grants unbiased surface extraction results. Analyst
influence is minimized, which leads to a very fast generation of risk areas. For a full
multi-layer Landsat data set the algorithm needs less than one minute (on a 512 MB
RAM computer with a double 2 GHz processor) for the calculations. Surface
extraction accuracies are as high as achieved with tedious manual classification
results (e.g. from maximum likelihood classifications). The above also applies for
the thermal algorithm. It can be transferred to other locations and so far works for
data from Landsat, ASTER and MODIS. The calculation time is about one hour
(on a 256MB RAM Laptop with a single 1.6 GHz processor) per full Landsat
thermal band.
Figure 19. Synergy of spectral test sequence (coal fire risk area delineation) and thermalalgorithm: rejection of false alarms outside the delineated risk areas.
4580 C. Kuenzer et al.
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7
(a) (b)
Figure 21. Newly detected coal fire in the valley east of the Yellow River about 27 kmsoutheast of Wuda city. The upper part of figure 19 shows a panchromatic subset of theLandsat ETM + image of 21 Sep 2002. The subset is about 15 km by 10 km large. Light blue:automatically extracted coal surfaces, yellow: coal fire risk areas, red: automatically detectedthermal anomalous clusters. s: solar induced thermal anomalies, i: industry inducedanomalies, c: confirmed newly detected coal fires. The orange arrow indicates the locationthe photographs in the lower part of figure 19 were taken.
Figure 20. Newly detected coal fire in the valleys of Hulusitai and Shitanjing about 30 kmnortheast of Ruqigou. The two parallel running valleys are shown on the right side in apanchromatic subset of the Landsat ETM + image of 21 Sep 2002. The subset is about 17 kmby 10 km. Light blue: automatically extracted coal surfaces, yellow: coal fire risk areas, red:automatically detected thermal anomalous clusters. s: solar induced thermal anomalies, i:industry induced anomalies, c: confirmed newly detected coal fires. On the left side you see acollapsed bedrock surface and a vent. Up to 340uC hot gasses are emitted through thesubsurface fire burning underneath.
Detecting unknown coal fires 4581
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7 The first discovery of formerly unknown coal fires using satellite remote sensing
data exclusively demonstrates that the combined approach of automated coal fire
area delineation and thermal anomaly detection holds the potential for observations
which go beyond the local and even regional scale. This supports the creation of an
operational detection and monitoring system for coal fires in large areas. It could be
demonstrated that the delineation of coal fire/risk areas can substantially reduce the
spatial area in which thermal investigations are necessary. On average less than 5%
of the ETM + input data sets were considered relevant for the detection of thermal
anomalies related to coal fires. For the full scene 129/33 (covering both main study
areas) only 1.5% of the entire area is considered of relevance for coal fire
investigations. To quantitatively evaluate the combination of the two algorithms
presented here, six scenes were processed. For the whole area covered by ETM +scene night-time scene (226/211), 451 thermal anomalies were extracted by the
thermal algorithm. 243 of these (ca. 50%) are located outside the delineated coal fire
risk areas. The remaining thermally anomalous clusters coincide well with the
known extent of coal fire areas. Within delineated risk areas only about 10%
overestimation occurs resulting from other heat sources. For the other daytime
scenes between 1334 and 10911 thermal anomalous clusters were extracted. These
numbers could be reduced by over 80% on average when applying the risk area
delineation for the exclusion of false alarms. However, in daytime data 50% of the
remaining anomalies do not stem from coal fires but from solar effects or other heat
sources. Thus, night-time and especially pre-dawn data is still the favourable choice
for all coal fire detection, including the automated method presented here. The
detectability of coal fires in Landsat, ASTER or MODIS data strongly depends on
the size and temperature of individual coal fires. While in Landsat night-time data
100% of all existing coal fires could be detected in the Ruqigou study area and 82%
could be detected in the Wuda area, in MODIS only 15% of the fires from Wuda
could be extracted from MODIS data. However, in Ruqigou, where fires are of large
spatial extent and relatively hot, even in MODIS data 80% of all fires could be
detected. We see the limitation of an operational coal fire monitoring not in
nowadays available methods, but in the spatial resolution of the data. If coal fires
lead to thermal anomalies which outpass the thermal variability of the general
background, they will be detected. However, if coal fires do not lead to a signal,
which can be noted in 60 m or 1 km data then it is even manually not possible to
extract the fire in the context of the overall image.
A major advantage of the algorithm for coal fire risk area delineation is the fact
that risk areas can be mapped, even if thermal anomalies have not yet been
developed. When new thermal anomalies are detected within delineated risk areas a
field check by local authorities can be triggered. Regular remote sensing based
monitoring of coal fire risk areas can enable a rapid response to newly ignited coal
fires at their initial stage. Coal fire fighting at this stage increases the probability to
control and extinguish the fire with relatively low effort. We plan to utilize the two
presented remote sensing algorithms for automated satellite data processing in
larger parts of the Chinese coal belt.
Acknowledgements
The authors thank the German Ministry for Education and Research (BMBF) for
the funding of the geo-scientific coal fire research initiative coordinated by the
4582 C. Kuenzer et al.
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7 German Aerospace Centre, DLR. Furthermore, we would like to thank all Chinese
and German project partners for their support and cooperation.
ReferencesBHATTACHARYA, A., REDDY, S. and MUKHERJEE, T.K., 1991, Multi-tir remote sensing data
analyses for coal fire mapping in Jharia coalfield of Bihar, India. In Proceedings of the
12th Asian Conference on Remote Sensing. Singapore, 30 Oct.–5 Nov. 1991, pp.
221–226.
BHATTACHARYA, A. and REDDY, C.S., 1994, Underground and surface coal mine fire
detection in India’s Jharia coalfield using airborne thermal infrared data. Asian
Pacific Remote Sensing Journal, 7, pp. 59–73.
CHEN, L., 1997, Subsidence Assessment in the Ruqigou Coalfield, Ningxia, China, using a
Geomorphological Approach. Master thesis at ITC, Enschede, Netherlands, pp. 130.
CIHLAR, J., 2000, Land cover mapping of large areas from satellites: status and research
priorities. International Journal of Remote Sensing, 21, No. 6&7, pp. 1093–1114.
ELLYETT, C.D. and FLEMING, A.W., 1974, Thermal infrared imagery of the Burning
Mountain coal fire. Remote Sensing of Environment, 3, No. 1, pp. 79–86.
FISHER, W. and KNUTH, W., 1968, Detection and delineation of subsurface coal fires by aerial
infrared scanning. Geological Society America, 115, pp. 67–68.
FRIEDL, M.A. and BRODLEY, C.E., 1996, Decision tree classification of land cover from
remotely sensed data. Remote Sensing of Environment, 61, pp. 399–409.
GANGOPADHYAY, P., 2003, Coalfire Detection and Monitoring in Wuda, North China – A
Multi-spectral and Multi-sensor TIR Approach. Master thesis at ITC. – Enschede,
Netherlands, 72 pp.
GAO, Y., MAS, J.F., MAATHUIS, B.H.P., ZHANG, X. and VAN DIJK, P., 2006, Comparison of
pixel-based and object-oriented image classification approaches—a case study in a
coal fire area, Wuda, Inner Mongolia, China. International Journal of Remote Sensing,
27, pp. 4093–4055.
GREENE, G.W., MOXHAM, R.M. and HARVEY, A.H., 1969, Aerial infrared surveys and
borehole temperature measurements of coal mine fires in Pennsylvania. In
Proceedings of the 6th Symposium on Remote Sensing of Environment, University of
Michigan, 13–16 October 1969 Institute of Science and Technology, University of
Michigan, USA, pp. 517–525.
GUAN, H., 1989, Applications of remote sensing techniques in coal geology. Acta Geologica
Sinica, 2, pp. 253–269.
GUPTA, R.P. and PRAKASH, A., 1998, Reflection aureoles associated with thermal anomalies
due to subsurface mine fires in the Jharia Coalfield, India. International Journal of
Remote Sensing, 19, pp. 2619–2622.
KNUTH, W.M., FISHER, W. JR. and STINGELIN, R.W., 1968, Detection, delineation and
monitoring of subsurface coal fires by aerial infrared scanning. Geographer HRB-Singer,
lnc., a subsidiary of the Singer Company, State College, Pennsylvania, USA, pp. 877–881.
KUENZER, C., 2005, Demarcating coal fire risk areas based on spectral test sequences and
partial unmixing using multi sensor remote sensing data. Ph.D. thesis, Technical
University Vienna, Austria, 199 pp.
KUENZER, C., STRUNZ, G., VOIGT, S. and WAGNER, W., 2004, Multitemporal Landcover
Investigations in a Semi-arid Mining Environment: Coal Fire Areas in Northern
China. In Proceedings of the Conference of the EARSeL Special Interest Group on
Land use and Land Cover, Dubrovnik, May 2004, pp. 31–37.
KUENZER, C. and VOIGT, S., 2003, Vegetationsdichte als moeglicher Indikator fuer
Kohlefloezbraende? Untersuchung mittels Fernerkundung und GIS. In Angewandte
Geographische Informationsverarbeitung XV. Beitraege zum 15. AGIT-Symposium
Salzburg 2003, J. Strobl, T. Blaschke and G. Griesebner (Eds), pp. 256–261
(Heidelberg, Germany: Wichmann, 2003).
Detecting unknown coal fires 4583
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7 KUENZER, C., ZHANG, J.A. and HIRNER, A., 2005, Multitemporal coal fire dynamics –
combining thermal remote sensing analysis and temperature field mappings to assess
coal fire development in Wuda coal mining area. In Proceedings of the 9th
International Symposium on Physical Measurements and Signatures in Remote
Sensing, ISPMSRS, 7–19 October 2005 Beijing, pp. 805–807.
KUENZER, C., ZHANG, J., TETZLAFF, A., VOIGT, S., VAN DIJK, P., WAGNER, W. and
MEHL, H., 2007, Uncontrolled coal fires and their environmental impacts:
Investigating two arid mining environments in north-central China. Applied
Geography, 27, pp. 42–62.
LI, J., VOIGT, V., KUENZER, C., YANG, B., ZHANG, J., ZHANG, Y., KONG, B. and ZHANG, S.,
2005, The progress in detecting coal fires on remote sensing – the first results of the joint
Sino-German research project in innovative technologies for exploration, extinction
and monitoring of coal fires in North China. In Proceedings of the Dragon Programme
Mid-Term Result Conference, 27.06.2005–01.07.2005, Santorini, Greece, pp. 3–12.
LI, D.X., 1985, Applied study of airborne remote sensing in Taiyuan coalfield, China. Internal
Report of the Remote sensing centre of the ministry of coal industry, Xian, China, pp.
3–10 (in Chinese).
MANSOR, S.B., CRACKNELL, A.P., SHILIN, B.V. and GORNYI, V.I., 1994, Monitoring of
underground coal fires using thermal infrared data. International Journal of Remote
Sensing, 15, pp. 1675–1685.
MOXHAM, R.M. and GREENE, G.W., 1967, Infrared surveys of coal mine fires in the
anthracite and bituminous fields, Pennsylvania, USA, Geological Survey Inter-
Agency Report BM-2, U.S. Bureau of mines., p. 67.
MUKHERJEE, T.K., BANDYOPADHYAY, T.K. and PANDE, S.K., 1991, Detection and
delineation of depth of subsurface coalmine fires based on an airborne multispectral
scanner survey in a part of the Jharia Coalfield, India. Photogrammetric Engineering
and Remote Sensing, 57, pp. 1203–1207.
PRAKASH, A. and GUPTA, R.P., 1998, Land-use mapping and change detection in a coal
mining area – a case study in the Jharia coalfield, India. International Journal of
Remote Sensing, 19, pp. 391–410.
PRAKASH, A., FIELDING, E.J., GENS, R., VAN GENDEREN, J.L. and EVANS, D.L., 2001, Data
fusion for investigating land subsidence and coalfire hazards in a coal mining area.
International Journal of Remote Sensing, 22, pp. 921–932.
PRAKASH, A., GUPTA, R.P. and SARAF, A.K., 1997, A Landsat TM based comparative study
of surface and subsurface fires in the Jharia Coalfield, India. International Journal of
Remote Sensing, 18, pp. 2463–2469.
PRAKASH, A., SARAF, A.K., GUPTA, R.P., DUTTA, M. and SUNDARAM, R.M., 1995, Surface
thermal anomalies with underground fires in Jharia coal mine, India. International
Journal of Remote Sensing, 16, pp. 2105–2109.
PUREVDORJ, T., TATEISHI, R., ISHIYAMA, T. and HONDA, Y., 1998, Relationships between
percent vegetation cover and vegetation indices. International Journal of Remote
Sensing, 19, pp. 3519–535.
RABCHEVSKY, G.A., 1972, Determination from available satellite and aircraft imagery of the
applicability of remote sensing techniques to the detection of fires burning in
abandoned coal mines and unmined coal deposits located in North Central Wyoming
and Southern Montana. In U.S. Department of Interior, Bureau of Mines, Prepared by
Allied Research Associates Inc., Contract No. SO 211087, Final Report No. 8G86-F,
21 pp.
RICHTER, R., 1998, Correction of satellite imagery over mountainous terrain. Applied Optics,
37, pp. 4004–4015.
SARAF, A.K., PRAKASH, A., SENGUPTA, S. and GUPTA, P., 1995, Landsat-TM data for
estimating ground temperature and depth of subsurface coal fire in the Jharia coal
field, India. International Journal of Remote Sensing, 16, pp. 2111–2124.
4584 C. Kuenzer et al.
Dow
nloa
ded
By:
[Sw
ets
Con
tent
Dis
tribu
tion]
At:
10:2
9 16
Nov
embe
r 200
7 SLAVECKI, R.J., 1964, Detection and location of subsurface coal fires. In Proceedings of the
third Symposium on Remote Sensing of Environment, 14–16 October 1964, Institute of
Science and Technology, University of Michigan, USA, pp. 537–547.
TETZLAFF, A., 2004, Coal fire quantification using Aster, ETM and Bird satellite instrument
data. Ph.D thesis at the Ludwig Maximilians University Munich, Germany, 155 pp.
VAN GENDEREN, J.L. and GUAN, H., 1997, Environmental monitoring of spontaneous
combustion in the North China coalfields. Enschede, Netherlands, 243 pp.
VOIGT, S., TETZLAFF, A., ZHANG, J., KUENZER, C., ZHUKOV, B., STRUNZ, G., OERTEL, D.,
ROTH, A., VAN DIJK, P. and MEHL, H., 2004, Integrating Satellite Remote Sensing
Techniques for Detection and Analysis of Uncontrolled Coal Seam Fires in North
China. International Journal of Coal Geology, No. 59, pp. 121–136.
YANG, B., CHEN, Y., LI, J., GONG, A., KUENZER, C. and ZHANG, J., 2005, Simple
normalization of multi-temporal thermal IR data and applied research on the
monitoring of typical coal fires in Northern China. In Proceedings of the Geoscience
and Remote Sensing Symposium, IGARSS 25–29 July 2005, Volume 8, pp. 5725–5728.
ZHANG, J., 1996, SWIR spectra of rocks in areas affected by coal fires, Xinjiang Autonomous
Region, P.R. of China. Master thesis at the ITC, Enschede, Netherlands, 74 pp.
ZHANG, J., 2004, Spatial and Statistical Analysis of Thermal Satellite Imagery for Extraction
of Coal Fire Related Anomalies. Ph.D. thesis at the Technical University Vienna,
Austria, 166 pp.
ZHANG, J., WAGNER, W., PRAKASH, A., MEHL, H. and VOIGT, S., 2004, Detecting Coal Fire
Using Remote Sensing Techniques. In International Journal of Remote Sensing, 25,
pp. 3193–3220.
ZHANG, X., 1998, Coal Fires in Northwest China – Detection, Monitoring, and Prediciton
Using Remote Sensing Data. PhD thesis at ITC Enschede, Netherlands, 235 pp.
ZHANG, X., ZHANG, J., KUENZER, C., VOIGT, S. and WAGNER, W., 2004, Capability
evaluation of 3–5mm and 8–12,5mm airborne thermal data for underground coalfire
detection. In International Journal of Remote Sensing, 25, pp. 2245–2258.
Detecting unknown coal fires 4585