PLEASE SCROLL DOWN FOR ARTICLE

This article was downloaded by:On: 26 May 2010Access details: Access Details: Free AccessPublisher Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

International Journal of Remote SensingPublication details, including instructions for authors and subscription information:http://www.informaworld.com/smpp/title~content=t713722504

Feature extraction for Darfur: geospatial applications in the documentationof human rights abusesJohn J. Sulika; Scott Edwardsb

a Department of Geography, The Florida State University, Tallahassee, FL, USA b AmnestyInternational USA, Washington, DC, USA

Online publication date: 20 May 2010

To cite this Article Sulik, John J. and Edwards, Scott(2010) 'Feature extraction for Darfur: geospatial applications in thedocumentation of human rights abuses', International Journal of Remote Sensing, 31: 10, 2521 — 2533To link to this Article: DOI: 10.1080/01431161003698369URL: http://dx.doi.org/10.1080/01431161003698369

Full terms and conditions of use: http://www.informaworld.com/terms-and-conditions-of-access.pdf

This article may be used for research, teaching and private study purposes. Any substantial orsystematic reproduction, re-distribution, re-selling, loan or sub-licensing, systematic supply ordistribution in any form to anyone is expressly forbidden.

The publisher does not give any warranty express or implied or make any representation that the contentswill be complete or accurate or up to date. The accuracy of any instructions, formulae and drug dosesshould be independently 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 directlyor indirectly in connection with or arising out of the use of this material.

Feature extraction for Darfur: geospatial applications in thedocumentation of human rights abuses

JOHN J. SULIK*† and SCOTT EDWARDS‡

†Department of Geography, The Florida State University, Tallahassee, FL 32306, USA

‡Amnesty International USA, Washington DC 20003, USA

(Received 15 February 2009; in final form 19 November 2009)

Geospatial technologies are rapidly becoming adopted by advocates of human

rights abuses, engaged in non-governmental monitoring. Limited available fund-

ing puts constraints on the amount of time staff can perform Geographic

Information System (GIS) and remote-sensing tasks. Therefore, semi-automated

techniques are forthcoming in order to facilitate data-analysis tasks aimed at

sharing information about violent conflict and human rights abuses. As a con-

tribution to these efforts, this paper details classification of a pre-conflict image,

binary partitioning of the classification results, application of morphological

filters, estimation of total number of pre-conflict structures and overlay of the

refined information onto the post-conflict image for damage assessment in Darfur,

Sudan. We present a novel application of geospatial technologies and image-

processing techniques aimed at expediting the dissemination of critical informa-

tion necessary to inform the public and policy makers of detailed multi-temporal

analyses of evidence of human rights abuses.

1. Introduction

The task of collecting information during an active conflict is one that frequently

stymies effective human rights monitoring and humanitarian planning. Internationalhuman rights organizations, such as Amnesty International and Human Rights

Watch, as well as local and regional organizations, have been collecting testimony

and ground-level observational data on abuses since the outbreak of hostilities in

Sudan’s westernmost provinces in early 2003. Despite the best efforts of these orga-

nizations, the immense size of Darfur, combined with persistent insecurity, has made

systematic and comprehensive documentation of potential abuses all but impossible.

Furthermore, and in response to escalating international outcry over the govern-

ment’s counterinsurgency tactics, the Government of Sudan (GoS) has severelylimited access to Darfur for international human rights organizations, media and

even humanitarian aid providers.

The effects of survey limitations are evident when considering the death toll of

Darfur as of late 2009, described as ranging from 200 000 to over 400 000 by disparate

studies in which the same operational definitions were used. Alternatively, the GoS

claims that no more than 9000 died as a result of the conflict. Moreover, the GoS

claims that departures from this estimate are exaggerations by ‘Western media and

NGOs’ (Non-governmental organizations; Reuters-Alertnet 2007). While this

*Corresponding author. Email: jsulik@fsu.edu

International Journal of Remote SensingISSN 0143-1161 print/ISSN 1366-5901 online # 2010 Taylor & Francis

http://www.tandf.co.uk/journalsDOI: 10.1080/01431161003698369

International Journal of Remote Sensing

Vol. 31, No. 10, 20 May 2010, 2521–2533

Downloaded At: 18:27 26 May 2010

particular claim is not widely viewed as credible, the inconsistent estimations raise a

pressing question: when information concerning human rights abuses is difficult to

attain in a particular area, what geospatial tools can be useful for advocates and

policy-makers to circumvent the data limitation?

Here, we explore the use of satellite image feature extraction in the case of Darfur,

with particular emphasis on the technical aspects of systematic data extraction, todetermine the number of extant dwellings in villages throughout the region. In doing

so, we present a method for using limited observation data of a semi-arid region that is

generally available to non-governmental organizations with reduced technical

and material capability. This method employs established image-processing

methods to overcome the inherent limitations of distinguishing indigenous housing

structures from their surrounding environment. Specifically, pixel-based operations

are employed to answer image-segmentation difficulties caused by spectral overlap

(figure 1). The goal of this study is to determine the feasibility of accurately quantify-ing the number of huts within a given village. Furthermore, we intend to demonstrate

the general applicability of such methods to the systematic collection of data for

purposes of humanitarian aid provision and/or abuse documentation and human

rights advocacy.

2. Background

2.1 Previous research

Many researchers have investigated the application of remotely sensed imagery to

humanitarian efforts (Bjorgo 2000, Brown et al. 2001, Giada et al. 2003). Giada et al.

Figure 1. Huts with conical roofs typically have a wide-range spectral response. The southeastportion of each hut often reflects high amounts of electromagnetic radiation with the northwestportion in shade.

2522 J. J. Sulik and S. Edwards

Downloaded At: 18:27 26 May 2010

(2003) automated tent extraction for a refugee camp in Tanzania, and as is the case in

many studies, the dwellings analysed were of a distinct shape and easily distinguished

from surrounding material. Giada’s study employed high-resolution IKONOS ima-

gery in which dwelling units were clearly identified; however, the authors decided that

classification of pixels into area classes was not as effective as object-oriented techni-ques or morphological operators (Giada et al. 2003). Unfortunately, specialized

software such as eCognition is prohibitively expensive for many researchers and

non-profit humanitarian organizations and requires extensive training beyond basic

image-processing skills, thus rendering it impractical for non-specialists to implement.

While the human rights application put forth here is novel, there are existing

Geographic Information System (GIS) methods and applications for humanitarian

purposes. For example, Prins (2008) performed change-detection analysis of Landsat

Enhanced Thematic Mapper Plus (ETMþ) data from a conflict-affected area ofDarfur in order to identify burned villages. Although useful for a medium-resolution

(30 m) approach, new problems become apparent when examining the burned villages

at a finer scale. Finer scale approaches often seek to isolate individual structures from

remotely sensed imagery, as opposed to an entire village. Mathematical morphology

has been applied in many remote-sensing studies, some of which are similar in applica-

tion to the one proposed here (Soille and Pesaresi 2002, Giada et al. 2003). The main

distinction between the application explored in this article and most of the existing

applications in the literature is the difficulty in segmenting the study area so thatfeatures of interest are isolated. Specifically, refugee camps are typically composed of

dwelling units that are easily distinguished from background features and have clear

outlines, whereas most of the villages inhabited by ethnic Africans in Darfur are made

of local materials and are not easily distinguishable from the surrounding area.

Darfur villages are comprised of compounds with lineaments that visibly intersect

dwelling units. Moreover, huts in Darfur villages do not exhibit a homogeneous

spectral response across their surface, a property exhibited by dwelling units in

other studies (Mason and Fraser 1998, Bjorgo 2000, Ruther et al. 2002). These areformidable barriers to data sampling (Lo 1995, Giada et al. 2003, De Laet et al. 2007).

Many studies often quantify structures in informal settlements through photogram-

metric techniques. These procedures rely on Digital Surface Models (DSMs) derived

from imagery acquired through air surveys (Mason and Fraser 1998, Ruther et al.

2002) to use an elevation component to enhance object identification; however, the

security situation in Darfur precludes this approach.

2.2 The conflict in Darfur

In the summer of 2003, following a series of defeats by armed opposition groups in

Darfur, the GoS began pouring military resources into Darfur and the surrounding

areas, heavily arming the Janjawid as a paramilitary force. These government-enabled

militias initiated a counter-insurgency campaign that relied on the destruction of com-

munities from which the rebels originated. As the number of groups taking up arms in

Darfur expanded, the targeting of civilians quickly spread to the rest of the country.

The government-enabled militias quickly gained the upper hand against the Darfurrebel movements, and by spring 2004, thousands of people, mostly civilians, had been

killed, and over a million people had been forcibly displaced. While over the course of

the conflict, the Darfur-armed opposition groups were responsible for many human

rights violations, the militia forces backed by GoS units systematically depopulated

Feature extraction for Darfur 2523

Downloaded At: 18:27 26 May 2010

large swaths of Darfur through the use of terror tactics. Anything that made life

possible was targeted for destruction; however, burnt domiciles are the most apparent

evidence in commercially available satellite sensor images.

3. Data

3.1 The case of Darfur: goals of and obstacles to automated data collection

Although geospatial technology is increasingly accessible to advocacy organizations,

there remain many encumbrances involved with remotely sensed imagery depicting

evidence of human rights abuses. First, the pre-conflict and post-conflict images are

rarely captured on anniversary dates. As a result, seasonal differences in soil moisture,

biomass, atmospheric variation and burn scars left after an attack complicate change

detection. The location of remotely observable conflict is rarely known prior to an

event and acquisition of a pre-conflict image can be subject to availability within thearchives of data providers such as DigitalGlobe. When reasonable pre- and post-

conflict images are accessible, the burden of analysis on resource-limited organiza-

tions is still substantial. Structures in a village number from the hundreds to well into

the thousands, making it infeasible to hand count them. This difficulty provided the

impetus to develop a method for automated and systematic assessment of dwellings.

Feature extraction within each village can only be successful if each free-standing

structure can be distinguished as a discrete object. If this criterion is met, the informa-

tion class that represents all of the huts in a scene can be vectorized and converted to ashapefile. The attribute table for each shapefile will have one record for each con-

tiguous group of pixels. The attribute table can then be appended with additional

fields and analysed within a GIS. A brief characterization of the study area will be useful

for method description. Groups of huts are usually surrounded by and touching grass

walls and fences with a similar spectral response. Moreover, most dwelling units of

sedentary farming communities have roofs and fences made out of local thatch. This

causes features of interest to be grouped in the same class as background information,

such as grass and shrubs. In addition, the fences are often so close to dwelling units thatthey appear to be connected when viewed from high-resolution satellite imagery. This

becomes problematic when trying to isolate dwelling units from fences and brush based

on digital number values of pixels. Further difficulty arises because the huts in each

village are small and have conical thatch roofs. These roofs cause light to scatter in

multiple directions, often resulting in heterogeneous pixel values for each hut. This

makes it difficult to spectrally identify each hut as a homogeneous object because up to

half of each roof falls in the shadow of the other half. To emphasize this, all pixel values

that are not representative of huts in the red band have a standard deviation of 25.96with a range of 255 (table 1). In contrast, the standard deviation of pixel values for huts

in the red band is 78.06 with a range of 255 (table 2).

Table 1. 2004 pixel statistics for background.

Basic statistics Minimum Maximum Mean Standard deviation

Band 1 1 160 100.79 13.07Band 2 1 190 132.70 17.89Band 3 1 255 192.76 25.96Band 4 1 246 168.92 22.08

2524 J. J. Sulik and S. Edwards

Downloaded At: 18:27 26 May 2010

3.2 Data sources

Standard Quickbird images of Jonjona village in North Darfur, Sudan were processed

using ENVI 4.2 and the MATLAB Image Processing Toolbox. This set of satellite

sensor imagery was kindly provided by the American Association for the Advancementof Science Geospatial Technology and Human Rights (GaTHR) program. The village

of Jonjona (figure 2) was reportedly attacked on 07 May 2006 when militias burned 16

houses within the village (American Association for the Advancement of Science 2009).

Due to image-availability restrictions, the pre-conflict image was captured on 07

December 2004 and the post-conflict image was captured on 23 February 2007; see

table 3 for characteristics of the images used in this study.

Digital image-processing procedures were performed using ENVI. Shapefiles

created with ENVI were opened in ArcMap for final analysis. Accuracy assessmentwas performed using manually digitized shapefiles (provided by the American

Association for the Advancement of Science’s Human Rights Program and

Table 2. 2004 pixel statistics for huts.

Basic statistics Minimum Maximum Mean Standard deviation

Band 1 1 165 55.15 47.16Band 2 1 199 65.17 58.23Band 3 1 255 91.51 78.06Band 4 1 242 70.87 70.88

Figure 2. False-colour infrared image of Jonjona village. Bands 4, 2 and 1 correspond towavelengths of 0.76–0.89 m (near-infrared), 0.52–0.60 m (green) and 0.45–0.52 m (blue),respectively.

Feature extraction for Darfur 2525

Downloaded At: 18:27 26 May 2010

Geospatial Technologies and Human Rights Project) indicating the number and

location of destroyed structures.

4. Methods

4.1 Classification procedures

The goal of this study is to determine the feasibility of accurately quantifying the

number of huts within a given village. As a consequence, the spatial extent of this

analysis was set by delineating a region of interest (ROI) around the village perimeter.

Furthermore, every feature in the scene besides a hut is considered background noise.Specifically, the classification goal was to sort all image pixels into two classes, one

representing huts and the other representing all other features in the scene.

Classification of the entire scene is unnecessary because the location and number of

huts in each image is the only information that is normally used for human rights

advocacy purposes. When comparing huts (table 2) to all other information content

(table 1) in the 07 December 2004 image, both categories have similar ranges of pixel

values across all spectral bands; however, the features of interest (huts) have larger

standard deviation values, owing mostly to the conical shape of hut roofs.Another impediment to classification was the Dynamic Range Adjustment (DRA)

option (table 3) that was applied by the data vendor. Unfortunately, images with

DRA are not recommended for spectral classification (DigitalGlobe 2009). To deal

with this, a simple non-parametric classification procedure was used to group pixels

between spectral bands. The ISODATA algorithm was chosen because of the non-

normal distribution of pixel values and poorly defined surface features within the

scene (Jensen 2007). The ISODATA algorithm treats the entire dataset as one cluster

and decomposes it into a number of natural spectral clusters after iterating throughthe dataset in a self-organizing way.

In order to better evaluate this procedure, the unsupervised ISODATA technique

was compared to two supervised classification techniques. Training data was col-

lected in order to prepare a Maximum Likelihood Classification. Training pixels were

collected across the entire study area, with over 300 pixels used for each of five classes.

This sample size was chosen in accordance with the principle that each sample should

minimally be 30 times the number of features in the dataset (Mather 2004).

Unfortunately, the histograms for each training set were extremely non-normal,and this was further aggravated by the large standard deviation of the pixels in the

training classes for huts and vegetation, thus causing the Maximum Likelihood

classifier to perform very poorly. A Support Vector Machine (SVM) was then used

because it is non-parametric and suitable for heterogeneous spectral classes. Instead of

relying on statistical criteria for class membership, SVM classifiers exploit geometric

criteria based on maximizing the margin between two classes (Melgani and Bruzzone

Table 3. Parameters of the Quickbird images used in this study.

Date View angleGround sampling

distance (m) Pan-sharpening DRA Map projection

07 December 2004 4.7 0.611 4 band Yes UTM zone 35N23 February 2007 – – 4 band Yes UTM zone 35N

2526 J. J. Sulik and S. Edwards

Downloaded At: 18:27 26 May 2010

2004). Although smaller training samples could have been suitable for the SVM (Foody

and Mathur 2004), the training sites used for the Maximum Likelihood classification

were reused for this analysis. The SVM classifier implemented in this study used the

pairwise classification strategy for multi-class classification. The parameters were

determined by trial analyses and a radial basis function kernel with a Gamma of 0.30was used in conjunction with a penalty parameter of 90. The results of the SVM

classification appear in table 4. Although the overall accuracy of the SVM classification

is rather good, the spectral class corresponding to huts and fencing still had multiple

errors and did not transfer well to the rest of the study. Because of this, only the

ISODATA results for the hut and fencing spectral class were further analysed.

ISODATA was initially set to iterate to seven clusters, and attempts were made to

recode the clusters into hut/non-hut classes, yet the results were unsatisfactory.

Repeated trial analyses determined that 80 clusters were optimal for recoding groupsof pixels into the two desired classification categories. It was necessary to break the

scene into many spectral clusters (using ISODATA) and then iteratively label the

resulting spectral classes, as ‘hut’ is an information class that is extremely hetero-

geneous and significantly overlaps with the spectral classes that occur for bare soil,

tree shadow, hut shadow, fencing and glare resulting from illumination. Each cluster

was interactively highlighted to determine correspondence to features of interest or

background information, ultimately being recoded into a binary image.

Binary images contain pixels that fall within two object sets: set A is the foreground(Boolean 1) and set B is the background (Boolean 0). In this study, village structures

and all other pixels with significant spectral overlap are in set A and all other pixels are

grouped into set B (figure 3). Generalization of pixel values into two groups has been

successfully used to quantify features within remotely sensed imagery (Glasbey et al.

1991, Laliberte and Ripple 2003) and prepares the image for transformations based

on set theory.

4.2 Removal of interfering features

Morphological operators are based on set theory and are similar to smoothing filters.

However, unlike convolution filters that act on spectral properties, morphological

filters modify the spatial properties of foreground pixels (set A) relative to back-

ground pixels (set B). Morphological transformations are applied through a struc-

tural element that dictates the connectivity (topology) of pixel groups from set A that

are allowed to pass through the filter. The connectivity within these structuralelements defines what information is retained from the original image (Serra and

Vincent 1992).

Table 4. Accuracy assessment of Support Vector Machine classification of entire village.

Class Commission (%) Omission (%)Producer

accuracy (%)User

accuracy (%)

Hut/fence 33.21 12.86 87.14 66.79Tree 0 8.26 91.74 100Grass 0.63 28.51 71.49 99.37Soil 1.77 0 100 98.23Overall accuracy ¼ 94.7784%Kappa coefficient ¼ 0.8998

Feature extraction for Darfur 2527

Downloaded At: 18:27 26 May 2010

This procedure has been extended to greyscale images (Sternberg 1986); however,

the current study only considers the application of morphological operators to binary

images. The two basic morphological operators are dilation and erosion and are

defined in equations (1) and (2), respectively,

X¯B ¼ xjBx ˙ X � ;f g (1)

and

X @B ¼ xjBx � Xf g (2)

for a binary image X and a structural element B where Bx ¼ bþ xjb 2 Bf g is the

translation of B at point x. Dilation operators fill holes smaller than the structuralelement and add pixels to object boundaries, whereas erosion operators remove pixels

from object boundaries. Additional operators can be implemented by combining

dilation and erosion. Closing operators are a combination of a dilation operator

followed by an erosion operator, essentially filling in gaps and removing narrow

features. Lastly, the opening operator is the reverse of a closing operator (Serra and

Vincent 1992). For an in-depth treatment of mathematical morphology, see Destival

(1986), Haralick et al. (1987) and Soille (2003).

Figure 3. Morphological filtering sequence with (a) huts and compound fencing as set A, (b) abinary image where set A is white and set B is black, (c) erosion of set A by set B and (d) dilationof set A in order to approximately expand the huts back to their original size.

2528 J. J. Sulik and S. Edwards

Downloaded At: 18:27 26 May 2010

The heterogeneous spectral response of image objects (see figure 2) underscores the need

for advanced processing techniques. For instance, similar colours are shared by tree

shadows and hut shadows, portions of hut roofs and bright patches of soil, and between

the compound fencing and hut roofs. Separation of these features based on pixel value

alone is intractable; therefore, the binary image was eroded then dilated. As detailed infigure 3, erosion is the dilation of set B and removes all features that do not correspond to

the parameters of the structural element, thus eliminating interfering features within the

remotely sensed image. In order to expand the huts back to their original shape, the

erosion operator was followed by the dilation operator, which erodes set B according to

the presence of foreground pixels within the structural element. This study uses 5 � 5 disc-

shaped structural elements because the features of interest are circular. Theoretically, any

stray pixels or (linear) features smaller than the size of the structural element are eroded.

An area threshold of 30 m was used to eliminate features larger than typical struc-tures, and the resulting file was overlain on the original image (figure 4). Therefore,

figure 4(b) represents a combination of the original image and the refined geographic

information that has been teased out of it in order to draw emphasis to dwelling units.

It is evident that overlay analysis makes it easier to count features in the pre-conflict

image by providing analysts with immediate visual cues regarding features of interest.

In addition, overlay analysis was also applied in this manner to the post-conflict

image to help discern destroyed structures (figure 4). Since the geographic informa-

tion from the pre-conflict image was carried over into the post-conflict image, thisoverlay procedure provides an elementary type of change detection. Specifically, this

form of multi-temporal analysis provides information on pre-existing structures that

have been obscured by burn scars and the passage of time, thus overcoming limita-

tions to interpreting change pairs via side-by-side comparison.

5. Results and discussion

The utility of combining geographic data with remotely sensed imagery to aid in post-disaster assessment has proved to be quite successful in other studies. However, this

paper differs in that the overlay procedure described herein uses generalized informa-

tion from the before image instead of overlaying ancillary geographic layers such as

population data.

Figure 4. Overlay analysis of extracted detail onto (a) pre-attack image and (b) post-attack image.

Feature extraction for Darfur 2529

Downloaded At: 18:27 26 May 2010

Given the inherent limitations in the dataset, the results are favourable (table 5). In

an effort to improve feature extraction accuracy, more advanced morphological

techniques were attempted. Specifically, morphological opening by reconstruction

(Soille 2003) was performed in an attempt to better preserve the shape of the man-

made structures; however, this was unsuccessful because much of the fencing aroundeach compound was contiguous with huts. As a result, this led to ‘over-reconstruc-

tion’. Specifically, the mask image allowed most of the fencing to be reconstructed

after the original binary image was eroded by a 5 � 5 disc-shaped structural element

(figure 5). The fact that this technique should logically improve feature boundaries yet

Table 5. Accuracy assessment of pre-conflict image feature extraction.

Identification accuracy Errors of commission Errors of omission

Proportion 284/424 86/424 52/424Accuracy (%) 66.9 20.3 12.8

Figure 5. Morphological reconstruction sequence showing (a) the binary image, (b) themorphological erosion with a 5 � 5 disc-shaped structuring element, (c) the opening byreconstruction that demonstrates how many smaller features were not reconstructed, yetfencing contours are still apparent, and (d) dilation of eroded image.

2530 J. J. Sulik and S. Edwards

Downloaded At: 18:27 26 May 2010

has the opposite effect, only places further emphasis on the difficulties imposed by the

spectral overlap between huts and compound fencing.

5.1 Quantitative accuracy evaluation

The accuracy of the hut extraction method was evaluated quantitatively by the use of

several accuracy indicators. Identification accuracy, error of omission and error of

commission were chosen because of their use in other feature extraction studies (Shi

and Zhu 2002, Zhu et al. 2005). Manually digitized vector files of huts were used for

the accuracy assessment. An omitted structure is one that was misclassified as back-

ground and a committed structure is background information that was misclassified

as a structure. Errors were calculated based on the number of huts correctly extracted

rather than estimated per-pixel (Congalton 1991, Giada et al. 2003). Errors of omis-sion resulted from tree shadows that have the same spectral response and are the same

size and shape as hut shadows, making it difficult to distinguish between the two.

Commission errors can be attributed to the difficulty in isolating tree shadows and are

caused by the same technical difficulties as the errors of omission. Improvement of

commission errors could possibly be attained with object-oriented methods; however,

it would be more time effective to manually remove committed structures.

Although not fully automated, the procedure presented here addresses the need of

multi-temporal quantitative analysis of pre/post-conflict imagery. This methodologysignificantly improves the feature extraction process considering the amount of

spectral overlap features of interest and background features (figure 2). The described

methodology of change observation via overlay analysis can be used by human rights

practitioners to aid in assessing the extent of damage that occurs in conflict areas,

considering that some settlements contain thousands of structures. Since visual

analysis of pre- and post-conflict imagery without image-processing techniques is

time-prohibitive, application of the proposed techniques can be used to expedite the

dissemination of critical information regarding evidence of human rights abuses.

6. Conclusion: current and future applications

There is clearly a need for rapid data collection and analysis methods with minimal

cost for human rights groups. The techniques presented would prove useful for most

semi-desert landscapes where it is difficult to separate individual settlement structures

from surrounding features. For instance, the circular hut with a thatch roof is

prevalent across African countries and can be observed from in desert climates,such as Chad, all the way to tropical climates found in places such as Burma. The

methods presented here describe the role remote-sensing tools and analysis techniques

can have in documenting the effects of militarized conflict on civilian centres and

allow for accurate assessment of the scope and scale of damage, with significantly

decreased costs associated with data collection. These techniques presented would

prove most useful for most semi-desert landscapes where it is difficult to separate

individual settlement structures from surrounding features.

The method presented in this paper exemplifies that the need for cost minimizationadds to the growing body of exemplars of the roles remote-sensing tools and analysis

techniques can have in documenting the effects of militarized conflict on civilian

centres. The steady decline in imagery costs, combined with the rapid growth of the

archives offers NGOs a valuable new tool. As demonstrated, however, mere collection

Feature extraction for Darfur 2531

Downloaded At: 18:27 26 May 2010

of this imagery without access to techniques that minimize the costs associated with

systematic analysis may make the successful use of the imagery still too costly.

Acknowledgements

The authors would like to thank the referees for the careful review and valuable

comments which provided insight that helped to improve the paper and Lars Bromley

for providing geospatial data, and to recognize Tingting Zhao, Elise Gornish, Juliette

Rousselot and William H. Moore for their input and valuable research assistance.

References

BJORGO, E., 2000, Using very high spatial resolution multispectral satellite sensor imagery

to monitor refugee camps. International Journal of Remote Sensing, 21, pp.

611–616.

AMERICAN ASSOCIATION FOR THE ADVANCEMENT OF SCIENCE, 2009, High-resolution satellite

imagery and the conflict in Chad and Sudan. Available online at: http://shr.aaas.org/

geotech/darfur/darfur.shtml (accessed 10 January 2009).

BROWN, V., JACQUIER, G., COULOMBIER, D., BALANDINE, S., BELANGER, F. and LEGROS, D., 2001,

Rapid assessment of population size by area sampling in disaster situations. Disasters,

25, pp. 164–171.

CONGALTON, R.G., 1991, A review of assessing the accuracy of classifications of remotely sensed

data. Remote Sensing of Environment, 37, pp. 35–46.

DE LAET, V., PAULISSEN, E. and WAELKENS, M., 2007, Methods for the extraction of archae-

ological features from very high-resolution Ikonos-2 remote sensing imagery, Hisar

(southwest Turkey). Journal of Archaeological Science, 34, pp. 830–841.

DESTIVAL, I., 1986, Mathematical morphology applied to remote sensing. Acta Astronautica, 13,

pp. 371–385.

DIGITALGLOBE, 2009, Quickbird Imagery Products Guide (Longmont, CO: DigitalGlobe).

FOODY, G.M. and MATHUR, A., 2004, Toward intelligent training of supervised image classifi-

cations: directing training data acquisition for SVM classification. Remote Sensing of

Environment, 93, pp. 107–117.

GIADA, S., DE GROEVE, T., EHRLICH, D. and SOILLE, P., 2003, Information extraction from very

high resolution satellite imagery over Lukole refugee camp, Tanzania. International

Journal of Remote Sensing, 24, pp. 4251–4266.

GLASBEY, C.A., HORGAN, G.W. and DARBYSHIRE, J.F., 1991, Image analysis and 3-dimensional

modeling of pores in soil aggregates. Journal of Soil Science, 42, pp. 479–486.

HARALICK, R.M., STERNBERG, S.R. and ZHUANG, X.H., 1987, Image analysis using mathema-

tical morphology. IEEE Transactions On Pattern Analysis And Machine Intelligence, 9,

pp. 532–550.

JENSEN, J.R., 2007, Remote Sensing of the Environment: An Earth Resource Perspective, 2nd edn

(Upper Saddle River, NJ: Prentice Hall).

LALIBERTE, A.S. and RIPPLE, W.J., 2003, Automated wildlife counts from remotely sensed

imagery. Wildlife Society Bulletin, 31, pp. 362–371.

LO, C.P., 1995, Automated population and dwelling unit estimation from high-resolution

satellite images – a GIS approach. International Journal of Remote Sensing, 16, pp.

17–34.

MASON, S.O. and FRASER, C.S., 1998, Image sources for informal settlement management.

Photogrammetric Record, 16, pp. 313–330.

MATHER, P.M., 2004, Computer Processing of Remotely Sensed Images: An Introduction, 4th edn

(Chichester, UK: Wiley).

2532 J. J. Sulik and S. Edwards

Downloaded At: 18:27 26 May 2010

MELGANI, F. and BRUZZONE, L., 2004, Classification of hyperspectral remote sensing images

with support vector machines. IEEE Transactions on Geoscience and Remote Sensing,

42, pp. 1778–1790.

PRINS, E., 2008, Use of low cost Landsat ETM to spot burnt villages in Darfur, Sudan.

International Journal of Remote Sensing, 29, pp. 1207–1214.

REUTERS-ALERTNET, 2007, Sudan’s Bashir defends record in Darfur. Available online at: http://

www.alertnet.org/thenews/newsdesk/L24114026.htm (accessed 23 February 2009).

RUTHER, H., MARTINE, H.M. and MTALO, E.G., 2002, Application of snakes and dynamic

programming optimisation technique in modeling of buildings in informal settlement

areas. ISPRS Journal of Photogrammetry and Remote Sensing, 56, pp. 269–282.

SERRA, J. and VINCENT, L., 1992, An overview of morphological filtering. Circuits Systems and

Signal Processing, 11, pp. 47–108.

SHI, W.Z. and ZHU, C.Q., 2002, The line segment match method for extracting road network

from high-resolution satellite images. IEEE Transactions on Geoscience and Remote

Sensing, 40, pp. 511–514.

SOILLE, P., 2003, Morphological Image Analysis: Principles and Applications (Berlin, Germany:

Springer-Verlag).

SOILLE, P. and PESARESI, M., 2002, Advances in mathematical morphology applied to

geoscience and remote sensing. IEEE Transactions on Geoscience and Remote Sensing,

40, pp. 2042–2055.

STERNBERG, S.R., 1986, Grayscale morphology. Computer Vision Graphics and Image

Processing, 35, pp. 333–355.

ZHU, C., SHI, W., PESARESI, M., LIU, L., CHEN, X. and KING, B., 2005, The recognition of road

network from high-resolution satellite remotely sensed data using image morphological

characteristics. International Journal of Remote Sensing, 26, pp. 5493–5508.

Feature extraction for Darfur 2533

Downloaded At: 18:27 26 May 2010