improving remote sensing ﬂood assessment using volunteered … · 2020-07-15 · the cryosphere...

Nat. Hazards Earth Syst. Sci., 13, 669–677, 2013www.nat-hazards-earth-syst-sci.net/13/669/2013/doi:10.5194/nhess-13-669-2013© Author(s) 2013. CC Attribution 3.0 License.

EGU Journal Logos (RGB)

Advances in Geosciences

Open A

ccess

Natural Hazards and Earth System

SciencesO

pen Access

Annales Geophysicae

Open A

ccess

Nonlinear Processes in Geophysics

Open A

ccess

Atmospheric Chemistry

and Physics

Open A

ccess

Atmospheric Chemistry

and Physics

Open A

ccess

Discussions

Atmospheric Measurement

Techniques

Open A

ccess

Atmospheric Measurement

Techniques

Open A

ccess

Discussions

Biogeosciences

Open A

ccess

Open A

ccess

BiogeosciencesDiscussions

Climate of the Past

Open A

ccess

Open A

ccess

Climate of the Past

Discussions

Earth System Dynamics

Open A

ccess

Open A

ccess

Earth System Dynamics

Discussions

GeoscientificInstrumentation

Methods andData Systems

Open A

ccess

GeoscientificInstrumentation

Methods andData Systems

Open A

ccess

Discussions

GeoscientificModel Development

Open A

ccess

Open A

ccess

GeoscientificModel Development

Discussions

Hydrology and Earth System

Sciences

Open A

ccess

Hydrology and Earth System

Sciences

Open A

ccess

Discussions

Ocean Science

Open A

ccess

Open A

ccess

Ocean ScienceDiscussions

Solid Earth

Open A

ccess

Open A

ccess

Solid EarthDiscussions

The Cryosphere

Open A

ccess

Open A

ccess

The CryosphereDiscussions

Natural Hazards and Earth System

Sciences

Open A

ccess

Discussions

Improving remote sensing flood assessment using volunteeredgeographical data

E. Schnebele and G. Cervone

Dept. of Geography and Geoinformation Science, George Mason University, 4400 University Drive, Fairfax, VA, USA

Correspondence to:E. Schnebele ([email protected]) and G. Cervone ([email protected])

Received: 6 September 2012 – Published in Nat. Hazards Earth Syst. Sci. Discuss.: –Revised: 19 November 2012 – Accepted: 23 January 2013 – Published: 19 March 2013

Abstract. A new methodology for the generation offlood hazard maps is presented fusing remote sensing andvolunteered geographical data. Water pixels are identifiedutilizing a machine learning classification of two Landsatremote sensing scenes, acquired before and during theflooding event as well as a digital elevation model paired withriver gage data. A statistical model computes the probabilityof flooded areas as a function of the number of adjacentpixels classified as water. Volunteered data obtained throughGoogle news, videos and photos are added to modify thecontour regions. It is shown that even a small amount ofvolunteered ground data can dramatically improve results.

1 Introduction

The ability to produce accurate and timely flood assessmentsis a critical safety tool for flood mitigation and response.Several methodologies have been developed to assess therisks associated with flooding by using ground measurementssuch as precipitation, water flow or level (e.g.Richter et al.,1998; Apel et al., 2006). Satellite remote sensing datahave been utilized for flood assessment because of theirhigh spatial resolution and capacity to provide informationfor areas of poor accessibility or lacking in groundmeasurements (Smith, 1997). High resolution satellite datais particularly useful for the spatial analysis of water pixels.When data before and after a flood event are available, itis possible to classify land cover change, and thus identifywhich areas are flooded.

The Landsat satellite program has been collecting dataabout the Earth and its environment since the 1970s, andhas been employed to monitor and mitigate the impacts of

flooding (Sanyal and Lu, 2004). The use of Landsat datafor flood assessment can be highly effective.Frazier andPage(2000) employed a supervised maximum-likelihoodclassification to map water bodies with Landsat ThematicMapper (TM), with an overall accuracy of over 97 %.Although effective for detecting water, the satellite’s orbitrevisit time can constrain data availability making it difficultto create a comprehensive time series of a flood event. Cloudand vegetative cover can obscure surface measurementswhen utilizing optical data, often resulting in partial coverageand incomplete flood assessment.

Numerous attempts have been proposed to overcome thelimitations of remote sensing data, often by supplementingthem with additional data to provide a more accurateand comprehensive flood assessment.Laura et al.(1990),Townsend and Walsh(1998) have proposed the use ofRADAR remote sensing data for the assessment of floods.RADAR has the unique advantage of penetrating throughcanopy and clouds, and can easily distinguish water bodiesfrom most other land cover types. However, RADAR data isnot widely available, and usually have limited swaths withlong revisit times. Efforts have been made toward increasingRADAR’s availability and accessibility.Wang et al.(2002)have proposed the integration of Landsat TM data with adigital elevation model (DEM) and river gage data to createa comprehensive assessment of flood depth under forest andcloud canopy. Although river gage data is usually sparse andnot universally available, especially in more remote areas;this methodology proved very robust and is routinely usedfor flood assessment. The research described in this paperis inspired by Wang’s methodology, where Landsat, DEM,and river gage data are used collectively in an attempt toimprove flood analysis. The main difference consists in the

Published by Copernicus Publications on behalf of the European Geosciences Union.

670 E. Schnebele and G. Cervone: Flood maps using satellite and volunteer data

fusing methodology employed in this study to integrate thedifferent data sources.

An emerging and quickly growing data source notyet fully utilized with respect to natural hazards isvolunteered geographic information (VGI) (Goodchild,2007). This general class of data, voluntarily contributed andmade available, contain temporal and spatial geographicalinformation. Data sources include pictures, videos, sounds,text messages, etc. Due to the spread of the internet to mobiledevices, an unprecedented and massive amount of grounddata have become available, often in real-time. Some dataare geolocated automatically, while others can be geolocatedby analyzing content.

Although volunteered data is often published withoutscientific intent, and usually carry little scientific merit,it is still possible to mine mission critical information.For example, during hurricane Katrina, geolocated picturesand videos searchable through Google provided earlyemergency response with ground-view information. Thesedata have been used during major events, with the capturein near real-time the evolution and impact of major hazards(De Longueville et al., 2009; Pultar et al., 2009; Heverin andZach, 2010; Vieweg et al., 2010; Acar and Muraki, 2011;Verma et al., 2011; Earle et al., 2012; Tyshchuk et al., 2012).

This work is based on a specific subset of this generalclass of data, namely photos, videos, and news. Volunteeredphotos about natural hazards have emerged as a data sourceduring crisis and hazardous events.Liu et al. (2008);Hyvarinen and Saltikoff(2010); McDougall (2011); Zhanget al.(2012) show how photos from Flickr have been used toderive local meteorological information, capture and recordthe physical features of an event, and identify and documentflood height.

Recently, volunteered data have been evaluated forestimating flood inundation depth and for mapping floodextent (Poser and Dransch, 2010; McDougall, 2011).These potentially valuable, real-time data have yet to beregularly applied in large scale disaster relief situationsfor multiple reasons, including difficulties of authenticationand confirmation, questions of quality and reliability,and difficulties associated with harvesting data fromheterogeneous and non-structured sources (Flanagin andMetzger, 2008; Schlieder and Yanenko, 2010; Tapia et al.,2011).

This paper proposes a new methodology that leveragesdata freely obtained from the internet to improve floodhazard estimation. Combining the high spatial resolutionand reliability of satellite imagery with the high temporalresolution of ground data takes advantage of the strengths ofboth data types while allowing for mutual data confirmation.

Despite the non-scientific nature of volunteeredinformation, the integration of these data with traditionaldata sources offers an opportunity to include new andadditional information in flood extent mapping. It isassumed that ground truth data is not available, and therefore

the quantitative analysis of the results discusses the changesintroduced by fusing the different data layers. The noveltyof this study is the development of a methodology that takesadvantage of “citizens as sensors”, as discussed inGoodchild(2007), to fuse observations culled from social media withsatellite and topographic data for flood assessment.

A case study is presented for the May 2011 flooding ofthe Mississippi River. This was one of the worst floods sincethe Great Flood of 1927. In Memphis, TN the MississippiRiver crested at 14.6 m, the highest crest since 1937,which caused the evacuation of approximately 1300 homes.The methodology was implemented using the R statisticalpackage1.

2 Data

2.1 Volunteered data

The data used in this study have been downloaded using theGoogle search engine through their photos, videos and newsportal. They included sources from Flickr, YouTube, WeatherUnderground, Wikipedia, and abc24.com. In particular,videos (n = 6) and photos (n = 8) from the first two weeksof May 2011 which documented the flooding were selected.A list of Memphis road closures on 12 May 2011 (n = 37)was collected from an on-line news source. Some of thedata contained geolocation information, while others weregeolocated using the Google API.

2.2 Remote sensing data

Full-resolution GeoTIFF Multispectral Landsat ETM+images for 2 January and 10 May 2011 are used. Thedata were downloaded from the USGS Hazards DataDistribution System (HDDS). Landsat data are comprisedof seven spectral bands: optical (0.45–0.52, 0.52–0.60, 0.63–0.69 µm), near-IR (0.77–0.90 µm), mid-IR (1.55–1.75, 2.09–2.35 µm) and thermal-IR (10.40–12.50 µm) with a spatialresolution of 30 m. The images were georeferenced to UTMcoordinates in ArcGIS and an area encompassing Memphisand its greater metropolitan area was selected at a scale of1:145 000.

2.3 Digital elevation model data

A USGS Seamless Data Warehouse DEM with a 30 mresolution was used in this study. The DEM is georeferencedto UTM coordinates in ArcGIS and exported at the same1:145 000 scale as the Landsat data (Fig.1).

2.4 Meteorological data

Meteorological data relative to maximum daily precipitationrate and total daily precipitation were obtained from the

1www.r-project.org

Nat. Hazards Earth Syst. Sci., 13, 669–677, 2013 www.nat-hazards-earth-syst-sci.net/13/669/2013/

www.r-project.org

E. Schnebele and G. Cervone: Flood maps using satellite and volunteer data 671

Figure 1: Digital Elevation Model for the region of study.

2.5. River Gage DataRiver gage data for the Mississippi River in Memphis2

was collected from the US Army Corps of EngineersRiverGages3 website. The data used for this study4

were collected from gage MS126 located at longitude:90.07667000 W, latitude: 35.12306000 N. Data were6

selected in elevation (meters) format so they couldeffectively be used in conjunction with the DEM.8

Figure 3 shows the height information for MS126for the entire year 2011. The acquisition time for the10

January and May Landsat data are indicated, and theycorrespond, respectively, to the almost minimum and12

maximum water heights for the entire year. The river gageheight information is paired with the DEM to derive the14

approximate flood extent.

3. Methodology16

3.1. OverviewThe proposed methodology is based on the data fusion18

of different layers generated from different data sources.Figure 4 illustrates this integration of multiple layers,20

which may have varying resolutions or sparse data. Theoutput is shown in the bottom most layer, where a22

flood hazard map is generated. The input consists ofdifferent layers generated using remote sensing data, DEM,24

ground information etc, as normally discussed in theliterature. The novelty of the proposed methodology, is26

the introduction of volunteered data as an additional layer,and their use in refining the hazard map. Therefore,28

although in this paper we used specific remote sensing and

3www.rivergages.com

Figure 2: Maximum daily precipitation rate and and accumulatedprecipitation for the period ranging from 1 April to 31 May 2011.

DEM data, the methodology is not limited by these data30

types and can easily be extended to integrate additionalor different sources. It is crucial for obtaining the32

most accurate measurements that the correct classificationmethodology is applied to each data type when creating34

the flood extent layer(s).The ground data used is not a verified “ground truth”,36

but can be utilized as reliable information to assess thepresence or absence of water in specific locations. It38

compensates for the potential miss classification of remotesensing data due to resolution, satellite orbit limitation,40

cloud cover, or data acquisition problems. Furthermore,the volume of the data alone as a function of time can be42

an indication of the geospatial rate of progression of theevent, and can help prioritize response to specific areas.44

The methodology consists of a three step process:

1. Identification of Flood Extent.46

2. Generation of flood hazard maps.3. Ground data integration.48

3.2. Identification of Flood ExtentDifferent methodologies can be used to identify the50

extent of water over the geographical region of interest.The goal of this step is to generate one or more maps52

using the input layers which identify regions where water isdetected. The task is method-independent, and it can use54

any method that is best suited for a particular combinationof data and location.56

In this article, two different methods are employed toidentify flood extent. The first involves the use of remote58

sensing data and machine learning classification, and thesecond the use of a DEM and river gage data.60

3.3. Generation of Flood Hazard MapsAfter one or more flood extent maps are generated, a62

flood hazard map is created by computing the probability

3

Fig. 1.Digital Elevation Model for the region of study.

NCEP CPC Morphed Precipitation Model (Joyce et al.,2004) and Weather Underground2 (WU), respectively.Figure2 shows the NCEP daily precipitation rate (bars) andthe WU accumulated precipitation (solid line) for the periodranging from 1 April to 31 May 2011. The acquisition timefor the May Landsat data is shown, and it occurs after theperiod of intense rainfall during the end of April. Thesemeteorological data are used to identify appropriate datesfor the remote sensing data. It is desirable that a scene isselected after a period of intense rainfall in order to identifythe maximum flood extent.

2.5 River gage data

River gage data for the Mississippi River in Memphiswere collected from the US Army Corps of EngineersRiverGages3 website. The data used for this studywere collected from gage MS126 located at longitude:90.07667000◦ W, latitude: 35.12306000◦ N. Data wereselected in elevation (m) format so they could effectively beused in conjunction with the DEM.

Figure 3 shows the height information for MS126 forthe entire year 2011. The acquisition time for the Januaryand May Landsat data are indicated, and they correspond,respectively, to the almost minimum and maximum waterheights for the entire year. The river gage height informationis paired with the DEM to derive the approximate floodextent.

2www.weatherunderground.com3www.rivergages.com

Figure 1: Digital Elevation Model for the region of study.

2.5. River Gage DataRiver gage data for the Mississippi River in Memphis2

was collected from the US Army Corps of EngineersRiverGages3 website. The data used for this study4

were collected from gage MS126 located at longitude:90.07667000 W, latitude: 35.12306000 N. Data were6

selected in elevation (meters) format so they couldeffectively be used in conjunction with the DEM.8

Figure 3 shows the height information for MS126for the entire year 2011. The acquisition time for the10

January and May Landsat data are indicated, and theycorrespond, respectively, to the almost minimum and12

maximum water heights for the entire year. The river gageheight information is paired with the DEM to derive the14

approximate flood extent.

3. Methodology16

3.1. OverviewThe proposed methodology is based on the data fusion18

of different layers generated from different data sources.Figure 4 illustrates this integration of multiple layers,20

which may have varying resolutions or sparse data. Theoutput is shown in the bottom most layer, where a22

flood hazard map is generated. The input consists ofdifferent layers generated using remote sensing data, DEM,24

ground information etc, as normally discussed in theliterature. The novelty of the proposed methodology, is26

the introduction of volunteered data as an additional layer,and their use in refining the hazard map. Therefore,28

although in this paper we used specific remote sensing and

3www.rivergages.com

Figure 2: Maximum daily precipitation rate and and accumulatedprecipitation for the period ranging from 1 April to 31 May 2011.

DEM data, the methodology is not limited by these data30

types and can easily be extended to integrate additionalor different sources. It is crucial for obtaining the32

most accurate measurements that the correct classificationmethodology is applied to each data type when creating34

the flood extent layer(s).The ground data used is not a verified “ground truth”,36

but can be utilized as reliable information to assess thepresence or absence of water in specific locations. It38

compensates for the potential miss classification of remotesensing data due to resolution, satellite orbit limitation,40

cloud cover, or data acquisition problems. Furthermore,the volume of the data alone as a function of time can be42

an indication of the geospatial rate of progression of theevent, and can help prioritize response to specific areas.44


1. Identification of Flood Extent.46

2. Generation of flood hazard maps.3. Ground data integration.48

3.2. Identification of Flood ExtentDifferent methodologies can be used to identify the50

extent of water over the geographical region of interest.The goal of this step is to generate one or more maps52

using the input layers which identify regions where water isdetected. The task is method-independent, and it can use54

any method that is best suited for a particular combinationof data and location.56

In this article, two different methods are employed toidentify flood extent. The first involves the use of remote58

sensing data and machine learning classification, and thesecond the use of a DEM and river gage data.60

3.3. Generation of Flood Hazard MapsAfter one or more flood extent maps are generated, a62

flood hazard map is created by computing the probability

3

Fig. 2. Maximum daily precipitation rate and and accumulatedprecipitation for the period ranging from 1 April to 31 May 2011.

Figure 3: Year 2011 water height profile for Mississippi River atMemphis, TN.

for each pixel to be flooded. This probability mapis generated by applying a kernel density smoothing2

operation over the 2D data, and then by normalizing theresult. Let (w1x1, w1x2, . . . , w1xn) be weighted samples4

drawn from a distribution with an unknown density f , thegoal is to estimate the shape of this function. The general6

kernel density estimator is

f(x) =1nh

n∑i=1

K(w(x− xi)

h) (1)8

where K is the kernel function, h is the bandwidth,and w is a user selected weighting scalar. The weight10

w describes the importance of a particular observation,or the confidence associated with the flood extent map.12

In this application, using a weighted kernel functionis paramount because ground observations cannot be14

considered “ground truth” proper, since volunteeredgeographical data carry intrinsic uncertainties due to their16

generally non-scientific nature. Therefore, using differentw values properly assigns levels of confidence to the various18

observations.The identification of w weights is problem specific and20

domain dependent, but most importantly, it is dependenton data quality. The w weights are used to include the22

concept of “significance” of the data in the algorithmand the analysis. It is assumed when working with24

such heterogeneous data, the information might varysignificantly, and therefore decisions should be made, when26

possible, using better data.Quantitative measures to define the w weights can be28

established. For example, when using satellite data thepixels along the center of the swath or those that are cloud30

free, are preferred in the analysis. Most satellite productshave a quality index associated with each pixel that can32

be used to set the appropriate w weight.

For volunteered data, the w weight may vary depending34

on the characteristics of the source. For example, thevolume of the data can be used to assign higher w36

weight to data with dense spatial coverage and numerousobservations. Higher w weight can also be dependent on38

the source itself. For example, observations published bylocal news are assumed to have been validated more than40

points volunteered anonymously. Finally, there is alsoa subjective component that can be taken into account,42

assigning different w weight to specific users or regions.The output of the kernel smoothing is a map with44

contour lines illustrating the probability that specificregions are flooded. The specific methodology used for46

Kernel smoothing and its R implementation used in thisstudy is described by Wand and Jones (1995).48

3.4. Ground Data IntegrationThe last step consists of modifying the flood hazard50

map by the integration of the ground data. The natureof this data is different from the data used to generate52

the previous flood extent maps. It is usually comprised ofsparse point data, identifying the presence or absence of54

flooding in a specific region.For this study, weight w values are assigned experi-56

mentally. They are first equally assigned and then ad-justed based on characteristics and confidence in the data58

source. By assuming the machine learning tree inductionand the DEM/river gage approach are equally adept at60

classifying water, their weight w values are kept constantwhile values of the volunteered ground data are adjusted.62

The flooded roads documented by local news sources areassumed more reliable than the sparse, point data of the64

videos and the pictures. Based on this assumption, theweights were set to 3 for the news data, and to 2 for the66

pictures and videos. Equal w values of 1 were assigned forboth the the DEM and Landsat data.68

4. Results

4.1. Flood Classification Using DEM and River Gage Data70

A DEM and river gage data are used to classify waterpixels for 2 January and 10 May 2011. Pixels in the DEM72

with a height less than or equal to the river gage height areset as water pixels. Specifically, heights of 56m and 70m74

are used for January and May, respectively. Figure 5a,cshow the areas identified as water for January and May76

dates, imposed over the DEM. The scale information isthe same as in Figure 1.78

4.2. Flood Classification Using Machine Learning TreeInduction80

Water pixels are identified in both the January andMay Landsat images by using a machine learning tree82

induction classifier. Ripley (2008) describes the generalrule induction methodology and its implementation in the84

R statistical package used in this study.

4

Fig. 3. Year 2011 water height profile for Mississippi River atMemphis, TN.

3 Methodology

3.1 Overview

The proposed methodology is based on the data fusionof different layers generated from different data sources.Figure4 illustrates this integration of multiple layers, whichmay have varying resolutions or sparse data. The output isshown in the bottom most layer, where a flood hazard mapis generated. The input consists of different layers generatedusing remote sensing data, DEM, ground information, etc.,as normally discussed in the literature. The novelty of theproposed methodology, is the introduction of volunteereddata as an additional layer, and their use in refining the

www.nat-hazards-earth-syst-sci.net/13/669/2013/ Nat. Hazards Earth Syst. Sci., 13, 669–677, 2013

www.weatherunderground.com

www.rivergages.com


Volunteered Data

DEM

Landsat Band 1

Landsat Band 2

Landsat Band 3

Landsat Band 4

Landsat Band 5

Landsat Band 6

Landsat Band 7

Flood Hazard Map

Figure 4: The proposed methodology fuses several layers to generatea flood hazard map.

For the machine learning classification, 4 controlareas of roughly the same size are identified, 2 over the2

Mississippi river as examples of water pixels and 2 overdifferent regions with no water pixels as counter-examples.4

Landsat multispectral data relative to these regions areused as training events by the decision tree classifier. The6

learned tree is then used to classify the remaining waterpixels in the scene. This process is repeated for both the8

January and May scene, and is illustrated in Figure 5b,d.About 1% of the total number of pixels are used10

as training pixels (events), and the remaining 99% areclassified according to the induction tree generated.12

4.3. Flood Hazard Maps and Ground Data IntegrationThe methodology described in Section 3.3 is employed14

to generate flood hazard maps using both the DEM andLandsat pixel classifications. The goal is to assign each16

pixel a probability of being part of the flooded area.Figure 6a,c show the probability contour lines for January18

and May, respectively.Additional data with ground information is then20

used to refine the January and May flood hazard maps(Figure 6b,d). The images show both the location and22

type of the ground information (Video, Photos, News),and the resulting probabilities when these data are taken24

into account. The data are imposed on both the Januaryand May hazard maps, although all ground data are26

collected from the May flood. The image generated forthe May flood, (Figure 6d), shows modifications to the28

flood hazard map after the incorporation of the grounddata. The ground data are also incorporated with the30

Januray (non-flood) image (Figure 6b) to illustrate howa preliminary hazard map could be generated if current32

satellite data are not available. In both instances, theaddition of supplemental information in the form of34

volunteered ground data alters the flood map by expandingthe area of possible inundation and by adjusting pixel36

values.The pixel classifications are summarized in Table 1 and38

by a histogram in Figure 7. As expected, when generallycomparing flood versus non-flood scenes as in Figure 6a,c,40

more pixels have a higher probability (60-100%) of beingflooded and less pixels have a lower probability (0-40%) of42

being flooded in the May (flooded) image as compared tothe January (non-flooded) image.44

When the ground data are incorporated into thehazard maps (Figure 6b,d), a spatial analysis shows46

noteworthy changes. The incorporation of ground datayields enhancements to the May flood hazard map48

(Figure 6d) which are evident by the progression ofcontour lines and reclassifications of pixels. Higher50

value contour lines, indicating a greater probability ofa region being flooded, progress toward the northeast,52

where the majority of ground information are located.Examining the differences between the two May scenes54

in Table 1, the percentage of pixels classified as having alow probability (0 - 20%) of being flooded as well as the56

pixels classified as having a high probability (80-100%)of flooding decreases or increases 6 percentage points,58

respectively, after the incorporation of ground data. Thesechanges illustrate that although both the DEM/river60

gage and Landsat classification techniques can be highlyaccurate in identifying flooded areas, the addition of real-62

time on the ground data, verifying the presence of waterin a specific area, can augment an inundation map.64

Applying the layer of ground data to the JanuaryHazard Map(Figure 6b) illustrates how a small amount66

of real-time volunteered ground data could be integratedwith an historical image to identify possible flooded68

regions. The number of non-water pixels (0-20%) isreduced by 7 percentage points and reclassified to higher70

probability classes (Table 1).Figure 6b,d shows while volunteered ground data does72

modify the flood hazard maps, the amount of modificationis limited by the spatial distribution of the ground data.74

The evolution of the contour lines, or areas of change, inboth images are restricted to regions where the volunteered76

data are located. This illustrates while the incorporationof volunteered ground data does affect a change in both78

flood hazard maps, the areas of change are controlled bythe quantity and distribution of the volunteered data.80

5

Fig. 4. The proposed methodology fuses several layers to generatea flood hazard map.

hazard map. Therefore, although in this paper we usedspecific remote sensing and DEM data, the methodology isnot limited by these data types and can easily be extendedto integrate additional or different sources. It is crucial forobtaining the most accurate measurements that the correctclassification methodology is applied to each data type whencreating the flood extent layer(s).

The ground data used is not a verified “ground truth”, butcan be utilized as reliable information to assess the presenceor absence of water in specific locations. It compensatesfor the potential misclassification of remote sensing datadue to resolution, satellite orbit limitation, cloud cover, ordata acquisition problems. Furthermore, the volume of thedata alone as a function of time can be an indication ofthe geospatial rate of progression of the event, and can helpprioritize response to specific areas.


1. Identification of flood extent.

2. Generation of flood hazard maps.

3. Ground data integration.

3.2 Identification of flood extent

Different methodologies can be used to identify the extentof water over the geographical region of interest. The goalof this step is to generate one or more maps using the input

layers which identify regions where water is detected. Thetask is method independent, and it can use any method that isbest suited for a particular combination of data and location.

In this article, two different methods are employed toidentify flood extent. The first involves the use of remotesensing data and machine learning classification, and thesecond involves the use of a DEM and river gage data.

3.3 Generation of flood hazard maps

After one or more flood extent maps are generated, a floodhazard map is created by computing the probability foreach pixel to be flooded. This probability map is generatedby applying a kernel density smoothing operation overthe 2-D data, and then by normalizing the result. Let(w1x1,w1x2, . . . ,w1xn) be weighted samples drawn froma distribution with an unknown densityf , the goal is toestimate the shape of this function. The general kerneldensity estimator is

f (x) =1

nh

n∑i=1

K

(w(x − xi)

h

), (1)

whereK is the kernel function,h is the bandwidth, andwis a user selected weighting scalar. The weightw describesthe importance of a particular observation, or the confidenceassociated with the flood extent map. In this application,using a weighted kernel function is paramount becauseground observations cannot be considered “ground truth”proper, since volunteered geographical data carry intrinsicuncertainties due to their generally non-scientific nature.Therefore, using different values ofw properly assigns levelsof confidence to the various observations.

The identification of weightw is problem specific anddomain dependent, but most importantly, it is dependent ondata quality. A weight is used to include the concept of“significance” of the data in the algorithm and the analysis. Itis assumed when working with such heterogeneous data, theinformation might vary significantly, and therefore decisionsshould be made, when possible, using better data.

Quantitative measures to define the weight can beestablished. For example, when using satellite data the pixelsalong the center of the swath or those that are cloud freeare preferred in the analysis. Most satellite products have aquality index associated with each pixel that can be used toset the appropriate weight.

For volunteered data, the weight may vary depending onthe characteristics of the source. For example, the volumeof the data can be used to assign higher weight to datawith dense spatial coverage and numerous observations.Higher weight can also be dependent on the source itself.For example, observations published by local news areassumed to have been validated more than points volunteeredanonymously. Finally, there is also a subjective componentthat can be taken into account, assigning different weight tospecific users or regions.



The output of the kernel smoothing is a map withcontour lines illustrating the probability that specific regionsare flooded. The specific methodology used for Kernelsmoothing and its R implementation used in this study aredescribed byWand and Jones(1995).

3.4 Ground data integration

The last step consists of modifying the flood hazard map bythe integration of the ground data. The nature of this data isdifferent from the data used to generate the previous floodextent maps. It is usually comprised of sparse point data,identifying the presence or absence of flooding in a specificregion.

For this study, weight values are assigned experimentally.They are first equally assigned and then adjusted basedon characteristics and confidence in the data source. Byassuming the machine learning tree induction and theDEM/river gage approach are equally adept at classifyingwater, their weight values are kept constant while values ofthe volunteered ground data are adjusted. The flooded roadsdocumented by local news sources are assumed more reliablethan the sparse, point data of the videos and the pictures.Based on this assumption, the weights were set to 3 for thenews data, and to 2 for the pictures and videos. Equal valuesof 1 were assigned for both the DEM and Landsat data.

4 Results

4.1 Flood classification using DEM and river gage data

A DEM and river gage data are used to classify water pixelsfor 2 January and 10 May 2011. Pixels in the DEM witha height less than or equal to the river gage height are setas water pixels. Specifically, heights of 56 m and 70 m areused for January and May, respectively. Figure5a, c show theareas identified as water for January and May dates, imposedover the DEM. The scale information is the same as in Fig.1.

4.2 Flood classification using machine learning treeinduction

Water pixels are identified in both the January and MayLandsat images by using a machine learning tree inductionclassifier.Ripley (2008) describes the general rule inductionmethodology and its implementation in the R statisticalpackage used in this study.

For the machine learning classification, 4 control areas ofroughly the same size are identified, 2 over the Mississippiriver as examples of water pixels and 2 over differentregions with no water pixels as counter-examples. Landsatmultispectral data relative to these regions are used astraining events by the decision tree classifier. The learnedtree is then used to classify the remaining water pixels in thescene. This process is repeated for both the January and Mayscene, and is illustrated in Fig.5b, d.

About 1 % of the total number of pixels are used astraining pixels (events), and the remaining 99 % are classifiedaccording to the induction tree generated.

4.3 Flood hazard maps and ground data integration

The methodology described in Sect. 3.3 is employed togenerate flood hazard maps using both the DEM and Landsatpixel classifications. The goal is to assign each pixel aprobability of being part of the flooded area. Figure6a, cshow the probability contour lines for January and May,respectively.

Additional data with ground information is then used torefine the January and May flood hazard maps (Fig.6b, d).The images show both the location and type of theground information (video, photos, news), and the resultingprobabilities when these data are taken into account. Thedata are imposed on both the January and May hazard maps,although all ground data are collected from the May flood.The image generated for the May flood, (Fig.6d), showsmodifications to the flood hazard map after the incorporationof the ground data. The ground data are also incorporatedwith the January (non-flood) image (Fig.6b) to illustratehow a preliminary hazard map could be generated if currentsatellite data are not available. In both instances, the additionof supplemental information in the form of volunteeredground data alters the flood map by expanding the area ofpossible inundation and by adjusting pixel values.

The pixel classifications are summarized in Table1 andby a histogram in Fig.7. As expected, when generallycomparing flood versus non-flood scenes as in Fig.6a, c,more pixels have a higher probability (60–100 %) of beingflooded and less pixels have a lower probability (0–40 %) ofbeing flooded in the May (flooded) image as compared to theJanuary (non-flooded) image.

When the ground data are incorporated into thehazard maps (Fig.6b, d), a spatial analysis showsnoteworthy changes. The incorporation of ground data yieldsenhancements to the May flood hazard map (Fig.6d)which are evident by the progression of contour linesand reclassifications of pixels. Higher value contour lines,indicating a greater probability of a region being flooded,progress toward the northeast, where the majority of groundinformation are located. Examining the differences betweenthe two May scenes in Table1, the percentage of pixelsclassified as having a low probability (0–20 %) of beingflooded as well as the pixels classified as having a highprobability (80–100 %) of flooding decreases or increases6 percentage points, respectively, after the incorporation ofground data. These changes illustrate that although both theDEM/river gage and Landsat classification techniques can behighly accurate in identifying flooded areas, the addition ofreal-time on the ground data, verifying the presence of waterin a specific area, can augment an inundation map.



Table 1.Number of pixels classified as water.

P(w) Jan Jan + Ground May May + Ground

0–20 % 126 567 (51 %) 108 932 (44 %) 114 104 (46 %) 99 368 (40 %)20–40 % 53 718 (21 %) 57 475 (23 %) 40 190 (16 %) 30 373 (12 %)40–60 % 21 136 (08 %) 27 601 (11 %) 21 235 (08 %) 26 997 (11 %)60–80 % 12 660 (05 %) 18 969 (08 %) 23 260 (09 %) 29 368 (12 %)

80–100 % 35 919 (14 %) 37 023 (15 %) 51 211 (20 %) 63 894 (26 %)

(a) Water Classification using DEM for Jan 2011 (b) Water Classification using Landsat for Jan 2011

(c) Water Classification using DEM for May 2011 (d) Water Classification using Landsat for May 2011

Fig. 5.Water pixel classification using the DEM (a andc) and Landsat (b andd) and for January (top) and May (bottom) data. The background(b andd) is from Landsat band 3.



(a) DEM + Landsat Jan 2011 (b) DEM + Landsat + Ground Jan 2011

(c) DEM + Landsat May 2011 (d) DEM + Landsat + Ground May 2011

Fig. 6. Flood hazard map indicating the probability of flood in percentage using DEM, Landsat, and ground data for January 2011 (a andb)and for May 2011 (c andd).

Applying the layer of ground data to the January HazardMap (Fig. 6b) illustrates how a small amount of real-time volunteered ground data could be integrated withan historical image to identify possible flooded regions.The number of non-water pixels (0–20 %) is reduced by7 percentage points and reclassified to higher probabilityclasses (Table1).

Figure 6b, d show while volunteered ground data doesmodify the flood hazard maps, the amount of modificationis limited by the spatial distribution of the ground data.The evolution of the contour lines, or areas of change, inboth images are restricted to regions where the volunteereddata are located. This illustrates while the incorporation ofvolunteered ground data does affect a change in both floodhazard maps, the areas of change are controlled by thequantity and distribution of the volunteered data.



P(w) Jan Jan+Ground May May+Ground

0-20% 126567 (51%) 108932 (44%) 114104 (46%) 99368 (40%)20-40% 53718 (21%) 57475 (23%) 40190 (16%) 30373 (12%)40-60% 21136 (08%) 27601 (11%) 21235 (08%) 26997 (11%)60-80% 12660 (05%) 18969 (08%) 23260 (09%) 29368 (12%)80-100% 35919 (14%) 37023 (15%) 51211 (20%) 63894 (26%)

Table 1: Number of pixels classified as water.

Figure 7: Number of pixels classified as water.

5. Conclusions

This paper presents a new methodology to fuse2

volunteered geographical data with remote sensing andDEM data for the creation of hazard maps. It is shown4

that even a small amount of ground data points can changethe flood assessment when compared to satellite and DEM6

data alone. The spatial distribution of the volunteereddata limits the areas where change is detected.8

The methodology is particularly useful when satellitedata is limited or of poor quality. Additionally, the ground10

data can add a time component which can help determinethe change occurring between remote sensing observations.12

References

Acar, A., Muraki, Y., 2011. Twitter for crisis communication: lessons14

learned from Japan’s tsunami disaster. International Journal ofWeb Based Communities 7 (3), 392–402.16

Apel, H., Thieken, A., Merz, B., Bloschl, G., 2006. A probabilisticmodelling system for assessing flood risks. Natural Hazards 38 (1),18

79–100.De Longueville, B., Smith, R., Luraschi, G., 2009. OMG, from20

here, I can see the flames!: A use case of mining location basedsocial networks to acquire spatio-temporal data on forest fires.22

In: Proceedings of the 2009 International Workshop on LocationBased Social Networks. ACM, pp. 73–80.24

Earle, P., Bowden, D., Guy, M., 2012. Twitter earthquake detection:earthquake monitoring in a social world. Annals of Geophysics26

54 (6).Flanagin, A., Metzger, M., 2008. The credibility of volunteered28

geographic information. GeoJournal 72 (3), 137–148.Frazier, P., Page, K., 2000. Water body detection and delineation30

with Landsat TM data. PE & RS- Photogrammetric Engineering& Remote Sensing 66 (12), 1461–1467.32

Goodchild, M., 2007. Citizens as sensors: the world of volunteeredgeography. GeoJournal 69 (4), 211–221.34

Heverin, T., Zach, L., on Information Systems for Crisis Response,I. C., Management, 2010. Microblogging for Crisis Communica-36

tion: Examination of Twitter Use in Response to a 2009 ViolentCrisis in the Seattle-Tacoma, Washington, Area. ISCRAM.38

Hyvarinen, O., Saltikoff, E., 2010. Social media as a source ofmeteorological observations. Monthly Weather Review 138 (8),40

3175–3184.Joyce, R., Janowiak, J., Arkin, P., Xie, P., 2004. Cmorph: A42

method that produces global precipitation estimates from passivemicrowave and infrared data at high spatial and temporal44

resolution. Journal of Hydrometeorology 5 (3), 487–503.Laura, L., Melack, J., David, S., 1990. Radar detection of flooding46

beneath the forest canopy: A review. International Journal ofRemote Sensing 11 (7), 1313–1325.48

Liu, S., Palen, L., Sutton, J., Hughes, A., Vieweg, S., 2008. In searchof the bigger picture: The emergent role of on-line photo sharing50

in times of disaster. Proceedings of ISCRAM 8.McDougall, K., 2011. Using volunteered information to map the52

queensland floods. Proc. Surveying & Spatial Sciences BiennialConference.54

Poser, K., Dransch, D., 2010. Volunteered geographic informationfor disaster management with application to rapid flood damage56

estimation. Geomatica 64 (1), 89–98.Pultar, E., Raubal, M., Cova, T., Goodchild, M., 2009. Dynamic58

GIS case studies: Wildfire evacuation and volunteered geographicinformation. Transactions in GIS 13 (1), 85–104.60

Richter, B., Baumgartner, J., Braun, D., Powell, J., 1998. Aspatial assessment of hydrologic alteration within a river network.62

Regulated Rivers: Research & Management 14 (4), 329–340.Ripley, B., 2008. Pattern recognition and neural networks. Cam-64

bridge Univ Pr.Sanyal, J., Lu, X., 2004. Application of remote sensing in flood66

management with special reference to monsoon asia: A review.Natural Hazards 33 (2), 283–301.68

Schlieder, C., Yanenko, O., 2010. Spatio-temporal proximity andsocial distance: a confirmation framework for social reporting.70

In: Proceedings of the 2nd ACM SIGSPATIAL InternationalWorkshop on Location Based Social Networks. ACM, pp. 60–67.72

Smith, L., 1997. Satellite remote sensing of river inundation area,stage, and discharge: A review. Hydrological processes 11 (10),74

1427–1439.Tapia, A., Bajpai, K., Jansen, B., Yen, J., Giles, L., 2011. Seeking76

the trustworthy tweet: Can microblogged data fit the informationneeds of disaster response and humanitarian relief organizations.78

In: Proceedings of the 8th International ISCRAM Conference. pp.1–10.80

Townsend, P., Walsh, S., 1998. Modeling floodplain inundationusing an integrated GIS with radar and optical remote sensing.82

Geomorphology 21 (3), 295–312.Tyshchuk, Y., Hui, C., Grabowski, M., Wallace, W., 2012. Social84

media and warning response impacts in extreme events: Resultsfrom a naturally occurring experiment. In: Proceedings of the86

45th Annual Hawaii International Conference on System Sciences(HICSS). IEEE, pp. 818–827.88

Verma, S., Vieweg, S., Corvey, W., Palen, L., Martin, J., Palmer, M.,Schram, A., Anderson, K., 2011. Natural language processing to90

the rescue? extracting situational awareness tweets during massemergency. Proc. ICWSM.92

Vieweg, S., Hughes, A., Starbird, K., Palen, L., 2010. Microbloggingduring two natural hazards events: what twitter may contribute94

to situational awareness. In: Proceedings of the 28th internationalconference on Human factors in computing systems. ACM, pp.96

1079–1088.Wand, M., Jones, M., 1995. Kernel smoothing. Vol. 60. Chapman &98

Hall/CRC.Wang, Y., Colby, J., Mulcahy, K., 2002. An efficient method for100

mapping flood extent in a coastal floodplain using Landsat TMand DEM data. International Journal of Remote Sensing 23 (18),102

3681–3696.

8

Fig. 7.Number of pixels classified as water.

5 Conclusions

This paper presents a new methodology fusing volunteeredgeographical data with remote sensing and DEM data for thecreation of hazard maps. It is shown that even a small amountof ground data points can change the flood assessment whencompared to satellite and DEM data alone. The spatialdistribution of the volunteered data limits the areas wherechange is detected.

The methodology is particularly useful when satellite datais limited or of poor quality. Additionally, the ground datacan add a time component which can help determine thechange occurring between remote sensing observations.

Acknowledgements.Work performed under this project has beenpartially supported by the US Department of Transportation award202717 (RITARS-12-H-GMU, CFDA).

The publication of this article was funded in part by the GeorgeMason University Libraries Open Access Publishing Fund.

Edited by: N. KerleReviewed by: three anonymous referees

References

Acar, A. and Muraki, Y.: Twitter for crisis communication: lessonslearned from Japan’s tsunami disaster, International Journal ofWeb Based Communities, 7, 392–402, 2011.

Apel, H., Thieken, A., Merz, B., and Bloschl, G.: A probabilisticmodelling system for assessing flood risks, Nat. Hazards, 38, 79–100, 2006.

De Longueville, B., Smith, R., and Luraschi, G.: “OMG, from here,I can see the flames!”: a use case of mining location basedsocial networks to acquire spatio-temporal data on forest fires,in: Proceedings of the 2009 International Workshop on LocationBased Social Networks, 73–80, 2009.

Earle, P., Bowden, D., and Guy, M.: Twitter earthquake detection:earthquake monitoring in a social world, Ann. Geophys., 54,708–715, 2012,http://www.ann-geophys.net/54/708/2012/.

Flanagin, A. and Metzger, M.: The credibility of volunteeredgeographic information, GeoJournal, 72, 137–148, 2008.

Frazier, P. and Page, K.: Water body detection and delineation withLandsat TM data, Photogramm. Eng. Rem. S., 66, 1461–1467,2000.

Goodchild, M.: Citizens as sensors: the world of volunteeredgeography, GeoJournal, 69, 211–221, 2007.

Heverin, T. and Zach, L.: Microblogging for crisis communication:examination of Twitter use in response to a 2009 violent crisisin Seattle-Tacoma, Washington area, in: Proceedings of theSeventh International Information Systems for Crisis Responseand Management Conference (ISCRAM), Seattle, Washington,2010.

Hyvarinen, O. and Saltikoff, E.: Social media as a source ofmeteorological observations, Mon. Weather Rev., 138, 3175–3184, 2010.

Joyce, R., Janowiak, J., Arkin, P., and Xie, P.: CMORPH: Amethod that produces global precipitation estimates from passivemicrowave and infrared data at high spatial and temporalresolution, J. Hydrometeorol., 5, 487–503, 2004.

Laura, L., Melack, J., and David, S.: Radar detection of floodingbeneath the forest canopy – A review, Int. J. Remote Sens., 11,1313–1325, 1990.

Liu, S., Palen, L., Sutton, J., Hughes, A., and Vieweg, S.: In searchof the bigger picture: the emergent role of on-line photo-sharingin times of disaster, in: Proceedings of the Information Systemsfor Crisis Response and Management Conference (ISCRAM),2008.

McDougall, K.: Using volunteered information to map theQueensland floods, in: Proceedings of the Surveying and SpatialSciences Biennial Conference, 2011.

Poser, K. and Dransch, D.: Volunteered geographic informationfor disaster management with application to rapid flood damageestimation, Geomatica, 64, 89–98, 2010.

Pultar, E., Raubal, M., Cova, T., and Goodchild, M.: DDynamicGIS case studies: wildfire evacuation and volunteered geographicinformation, Transactions in GIS,13, 85–104, 2009.

Richter, B., Baumgartner, J., Braun, D., and Powell, J.: A spatialassessment of hydrologic alteration within a river network,Regul. River., 14, 329–340, 1998.

Ripley, B.: Pattern recognition and neural networks, CambridgeUniversity Press, 2008.

Sanyal, J. and Lu, X.: Application of remote sensing in floodmanagement with special reference to monsoon Asia: a review,Nat. Hazards, 33, 283–301, 2004.

Schlieder, C. and Yanenko, O.: Spatio-temporal proximity andsocial distance: a confirmation framework for social reporting,in: Proceedings of the 2nd ACM SIGSPATIAL InternationalWorkshop on Location Based Social Networks, ACM, 60–67,2010.


http://www.ann-geophys.net/54/708/2012/


Smith, L.: Satellite remote sensing of river inundation area, stage,and discharge: A review, Hydrol. Process., 11, 1427–1439, 1997.

Tapia, A., Bajpai, K., Jansen, B., Yen, J., and Giles, L.:Seeking the trustworthy Tweet: can microblogged data fitthe information needs of disaster response and humanitarianrelief organizations, in: Proceedings of the 8th InternationalConference on Information Systems for Crisis Response andManagement (ISCRAM), 10 pp., 2011.

Townsend, P. and Walsh, S.: Modeling floodplain inundationusing an integrated GIS with radar and optical remote sensing,Geomorphology, 21, 295–312, 1998.

Tyshchuk, Y., Hui, C., Grabowski, M., and Wallace, W.: Socialmedia and warning response impacts in extreme events: resultsfrom a naturally occurring experiment, in: Proceedings ofthe 45th Annual Hawaii International Conference on SystemSciences (HICSS), IEEE, 818–827, 2012.

Verma, S., Vieweg, S., Corvey, W., Palen, L., Martin, J., Palmer,M., Schram, A., and Anderson, K.: Natural language processingto the rescue? Extracting “situational awareness” Tweets duringmass emergency, in: Proceedings of the ACM 2012 conferenceon Computer Supported Cooperative Work Companion, 227–230, 2011.

Vieweg, S., Hughes, A., Starbird, K., and Palen, L.: Microbloggingduring two natural hazards events: what twitter may contribute tosituational awareness, in: Proceedings of the 28th internationalConference on Human Factors in Computing Systems, ACMSIGCHI, 1079–1088, 2010.

Wand, M. and Jones, M.: Kernel smoothing, Vol. 60, Chapman &Hall/CRC, 1995.

Wang, Y., Colby, J., and Mulcahy, K.: An efficient method formapping flood extent in a coastal oodplain using Landsat TMand DEM data, Int. J. Remote Sens., 23, 3681–3696, 2002.

Zhang, H., Korayem, M., Crandall, D., and LeBuhn, G.: Miningphoto-sharing websites to study ecological phenomena, in:Proceedings of the 21st international conference on World WideWeb, 749–758, 2012.


improving remote sensing ﬂood assessment using volunteered … · 2020-07-15 · the cryosphere...

Documents