Landsat-based early warning system to detect the destruction of villages in Darfur, Sudan
Introduction
In response to the crimes committed against humanity during the Second World War, the global community, through the newly formed United Nations (U.N.), drafted the Universal Declaration of Human Rights to identify and establish a mandate to protect universal human rights. Today the U.N., along with other governmental and non-governmental organizations (NGOs), conducts imagery-based human rights monitoring campaigns to document any governmental and non-state transgressions and abuses of human rights and the failures of the state to protect these rights (Marx and Goward, 2013, OHCHR [Office of the High Commissioner for Human Rights], 2001). These campaigns are systematic, long-term efforts over a large region, using tens or hundreds of satellites images to detect phenomena associated with violations of human rights and international humanitarian law. Satellite observations have remedied biases in human ecology studies related to difficulty of accessing many remote areas and subsequent preferential sampling of sites in proximity to roads and reliance on single snapshot observations in characterizing those sites; however, they have only partially removed another major limitation of human ecology research — heavy reliance on visual description of patterns (Turner, 2003).
Although very high resolution (VHR) (10 meter and higher spatial resolution) imagery has been used in since the 1990s to document specific human rights events such as the 1995 massacre in Srebrenica, Bosnia (NYT, 1995), one of the first examples of a systematic, imagery-based project to document and publicize human rights violations over a long period of time in a large region was conducted by the U.S. Department of State's Office of the Geographer and Global Issues' Humanitarian Information Unit (HIU). The HIU compiled information based on hundreds of VHR satellite images and published a map which showed the widespread destruction of villages in Darfur, Sudan (HIU, 2004). This information was passed to the U.N. and NGOs to assist in their efforts to monitor and document human rights violations in Darfur. By 2011 one especially well-funded human rights monitoring project, the Satellite Sentinel Project, had demonstrated the ability to purchase imagery from a constellation of VHR sensors and publicize the imagery from a location with a reported human rights violation in as little as 24 h (SatSentinel, 2011). Many other groups concerned with the human rights violations adopted similar analyst-intensive approaches to mostly visual interpretation of VHR imagery to monitor the conflict in Darfur and many other regions of the world (AAAS, 2013).
Although the goals, production timeline, and publically released products of these human rights monitoring campaigns have been refined over the years, there are only a few examples of their methods evolving (Sulik and Edwards, 2010, Wolfinbarger and Drake, 2012) limiting the growth of remote sensing in the human rights community. Relying on manual analysis of VHR imagery over large areas and over long periods of time, these campaigns are cost prohibitive to all but the most well-funded monitoring efforts (Pisano, 2011). In the past, this work has been limited that it requires eyewitness reports, either from victims or international observers, before an organization can order satellite images to gather evidence of suspected violations. In the future social media, through mobile devices and internet penetration in many regions, promises to reduce the time gap between a reported human rights violation and an organization ordering imagery of that location.
Satellite and data-based early warning systems have been used since the 1970s for crop forecasts, severe storm prediction, and land use planning (Estes et al., 1980), but the development of early warning systems based on remotely sensed data to automatically alert users of a condition is relatively recent. One early warning system is the Drought Monitor, which uses data from the Advanced Very High Resolution Radiometer (AVHRR), in addition to other inputs, to provide weekly maps identifying where the emerging, persisting, and subsiding conditions of drought are located (Wilhite & Svoboda, 2000). The United States Agency for International Development (USAID) uses the Famine Early Warning System Network (FEWS NET) and takes this system one step further, incorporating AVHRR NDVI and meteorological satellites' rainfall data with other field-collected data to identify problems in the food supply system that could lead to famine or other food-insecure conditions (www.fews.net). FEWS NET products are available on a 10-day time period. In addition to these systems, the Webfire Mapper is part of the U.N.'s Foreign Agricultural Organization's Global Fire Information Management System (GFIMS). The Webfire Mapper was developed by the University of Maryland and supported by NASA, and it provides near real-time information on active fires worldwide, detected by the Moderate Resolution Imaging Spectroradiometer (MODIS) (Davies et al., 2009).
Because existing early warning systems rely on daily image acquisition, their value is limited to applications developed from coarse resolution (greater than 100 meter pixel size) imagery. The large pixel size of coarse resolution imagery makes definitive identification of most phenomena associated with human rights violations in this dryland ecosystem extremely challenging. Villages in arid environments can be relatively small (100 meter in diameter) and are constructed largely from local material, such as mud brick and roofs made from dry plant material, and show little contrast with the natural vegetation background in coarse resolution imagery. These differences have however been detected when the near infrared (NIR) band in VHR is utilized (Sulik & Edwards, 2010).
Research into applications from moderate resolution imagery (10–100 meter) has provided promising results, though no fully operational methods to monitor areas at risk of human rights violations have yet been developed. Prins (2008) demonstrated that Landsat Enhanced Thematic Mapper Plus (ETM +) could detect the destruction of villages in Darfur on an annual basis based on a visual analysis in the drop of albedo and Bromley (2010) links the Moderate Resolution Imaging Spectroradiometer (MODIS) sensor, which detects fire, with eyewitness reports of violence in Darfur. Landsat, Indian Remote-Sensing Satellite (IRS) and Système Pour l'Observation de la Terre (SPOT) imagery was used in 1999 to identify war-induced agricultural abandonment in Kosovo over a two-year period (Terres et al., 1999). Other studies have followed, demonstrating the ability of Landsat Thematic Mapper (TM) to detect agricultural abandonment in Bosnia in response to armed conflict (Witmer, 2008). Schimmer (2008) used coarse resolution MODIS and moderate resolution SPOT imagery to track an increase in vegetation cover and vigor in Darfur, which he linked to the displacement of people and their livestock. He noted that the spatial and temporal tracking of population displacement could provide important evidence to the scale and systematic nature of the violence—essential evidence in genocide trials.
Although previous studies have shown promising results, these methods are more suitable to scientific research than operational monitoring due to the considerable lag in time between the impact of the armed conflict on population and its identification in satellite imagery. In Preventing Genocide: A Blueprint for U.S. Policymakers, former U.S. Secretary of State Madeleine Albright writes that “at its most basic level, early warning means getting critical information to policymakers in time for them to take effective preventive action” (2008).
The strategic data acquisition plan for Landsat missions provides a suitable data source to serve as a prototype for development of such a warning system (Goward et al., 2006). The 16-day repeat cycle from Landsat 7 allowed us to collect an archive of images over Darfur between 1999 and 2011. While a combination of two Landsat satellites would have provided a better return frequency of 8 days, and thus greater opportunities for monitoring, this was not possible over Darfur region, as Landsat 5 is functioning in a very limited capacity and does not cover Darfur region at present and Landsat 7 has a scan line corrector (SLC) malfunction since mid-2003 leaving only 75% of each scene usable (http://landsat.gsfc.nasa.gov/about/landsat7.html). Subsequently, this project uses ETM + as a prototype to develop such a warning system and to test its abilities. It may then be operationally deployable now that Landsat Data Continuity Mission (LDCM) is online as of February 2013.
Methods applied within an early warning system require an economically viable combination of frequent observations of the affected area, as well as an appropriate spatial resolution and spectral range for detecting the footprint of the phenomena associated with human rights violations. Currently, none of the available satellite systems meet these requirements, although, when functioning, the combination of Landsat 5 and 7 did provide free weekly observations of the area with data collected across visible, NIR, and short-wave infrared (SWIR) spectrum of electromagnetic radiation with limited resolving power of individual objects on the ground. In contrast, the price of individual VHR images, which frequently have limitations in spectral range of observations, makes many operational monitoring applications economically unfeasible. However, only VHR imagery allows for definitive identification of individual households and their condition in the Darfur region. Therefore, an early warning system for monitoring impacts of an armed conflict on population in Darfur requires a coordinated effort of moderate resolution imagery for early-stage possible identification and VHR imagery for verification.
In this paper we introduce a methodology for an early warning system using ETM + that is designed to provide automated detection of the destruction of villages in arid environments. This remote sensing algorithm capitalizes on Landsat program's historical archive, radiometric stability, consistent calibration, and systematic observations by constructing a historic spectral baseline for each village in the study area (Markham et al., 2004). The application of the algorithm to the archived or operationally acquired Landsat and Landsat-like imagery identifies areas of high likelihood for village destruction in space-time and provides specifications for VHR image acquisition and analysis to verify impact and quantify the extent of damage at the individual household level.
Section snippets
Study area
In the late 1980s and 1990s, the Sudanese state of Darfur experienced clashes from both inter-tribal conflicts and armed insurrection by rebel groups. Beginning in 2003, the violence significantly escalated as government-supported militia groups, and later Sudanese military forces, attacked and destroyed thousands of villages. By September 2005 over 2 million people had fled the rural areas of Darfur to camps and the larger towns, and another 200,000 had sought refuge in neighboring Chad (
Methodology
The input data for the algorithm includes Landsat surface reflectance data and the Humanitarian Information Unit's database of villages in Darfur (HIU, 2010). The methodology is presented in 3 parts: 1) image processing and village delineation, 2) algorithm flow, and 3) evaluation of 16 bands/indices derived from ETM +.
Validation dataset and methodology
The validation reference base used to test the algorithm's performance on these 197 villages is the U.S. Department of State's Office of the Geographer and Global Issues' HIU database (HIU, 2010). This database is a compilation of information based on hundreds of VHR satellite images that lists the location of villages in Darfur, Sudan and their annual status as either damaged, destroyed, or no damage. A destroyed village is defined as confirmed evidence of complete destruction of the village.
Results
The use of ETM + band 4 in the algorithm correctly detected 84% of the villages identified as destroyed by the HIU database in 2004. The algorithm incorrectly detected 14% of the control villages as being destroyed. Increasing the significance level to 0.0005 reduces the omission rate, but it also significantly increases the number of villages incorrectly detected as destroyed. While there is limited VHR imagery of this area in 2004 and subsequent years, a survey of available VHR imagery from
Conclusion
Currently few organizations are able to conduct human rights monitoring campaigns because existing methods are prohibitively costly and labor-intensive. The presented algorithm provides an approach that reduces the cost of human rights monitoring campaigns in arid regions by focusing the purchase of VHR imagery and analysis to areas that have been alerted by moderate resolution imagery sensors. This does not eliminate the costs however. While the moderate resolution imagery and imagery
References (42)
- et al.
Impacts of remote sensing on US geography
Remote Sensing of Environment
(1980) - et al.
Complementarity of resourcesat-1 AWiFS and Landsat TM/ETM + sensors
Remote Sensing of Environment
(2012) - et al.
Locating and estimating the areal extent of wildfires in Alaskan boreal forests using multiple-season AVHRR NDVI composite data
Remote Sensing of Environment
(1995) - et al.
Trajectory-based change detection for automated characterization of forest disturbance dynamics
Remote Sensing of Environment
(2007) - et al.
Regionally adaptable dNBR-based algorithm for burned area mapping from MODIS data
Remote Sensing of Environment
(2007) Red and photographic infrared linear combinations for monitoring vegetation
Remote Sensing of Environment
(1979)AAAS scientific responsibility, human rights and law program: Geospatial technologies and human rights
(2013)Relating violence to MODIS fire detections in Darfur, Sudan
International Journal of Remote Sensing
(2010)- et al.
Review ArticleDigital change detection methods in ecosystem monitoring: A review
International Journal of Remote Sensing
(2004) - et al.
Fire information for resource management system: Archiving and distributing MODIS active fire data
IEEE Transactions on Geoscience and Remote Sensing
(2009)
GLOBCPVER: A 300m global land cover product for 2005 using ENVISAT MERIS Time Series
Historical record of Landsat global coverage: Mission operations, NSLRSDA, and international cooperator stations
Photogrammetric Engineering and Remote Sensing
Sudan (Darfur)–Chad border region confirmed damaged and destroyed villages
Darfur, Sudan: Confirmed damaged and destroyed villages, February 2003–December 2009
Derivation of a tasselled cap transformation based on Landsat 7 at-satellite reflectance
International Journal of Remote Sensing
Improving the global precipitation record: GPCP version 2.1
Geophysical Research Letters
Characterization of the Landsat-7 ETM automated cloud-cover assessment (ACCA) algorithm
Photogrammetric Engineering and Remote Sensing
Measuring and remote sensing of burn severity
Landsat-7 ETM + on-orbit reflective-band radiometric stability and absolute calibration
IEEE Transactions on Geoscience and Remote Sensing
Remote sensing in human rights and international humanitarian law monitoring: concepts and methods
Geographical Review
Cited by (16)
Hybrid U-Net: Semantic segmentation of high-resolution satellite images to detect war destruction
2022, Machine Learning with ApplicationsDetecting village burnings with high-cadence smallsats: A case-study in the Rakhine State of Myanmar
2019, Remote Sensing Applications: Society and EnvironmentCitation Excerpt :The method pairs a destroyed village documented in the ground reference dataset with another nearby but intact village displaying a similar phenology. This ‘pairing’ of villages is unusual because typically all villages in a conflict area are destroyed (Marx and Loboda, 2013). However, in this case the segregated Rohingya villages are typically next to an intact Rahkine village (Human Rights Watch, 2017).
Detecting urban destruction in Syria: A Landsat-based approach
2016, Remote Sensing Applications: Society and EnvironmentCitation Excerpt :Because the difference in surface reflectance between pre-destruction and post-destruction pixels is very slight, traditional methods such as band threshold or image differencing were insufficient in detecting the destruction. Marx and Loboda (2013) demonstrated that a small change to a group of pixels could be detected by comparing each new group of pixels to a historic baseline; however, when this analysis was conducted for individual pixels in Syria, it was found to be insufficient due to the small difference in pre- and post-destruction surface reflectance and the intra-annual changes in those pixels. In response, an intra-annual curve was calculated for every pixel in the study area based off images collected during the six-year baseline.
Using multi-temporal Landsat imagery and linear mixed models for assessing water quality parameters in Río Tercero reservoir (Argentina)
2015, Remote Sensing of EnvironmentCitation Excerpt :Landsat 5 has problems with its solar array drive which have affected data availability (Furby & Wu, 2009). In addition, as this satellite has exceeded its design life, fuel is not enough to maintain operations through time, and therefore no longer has global coverage (Marx & Loboda, 2013; Wulder, White, Masek, Dwyer, & Roy, 2011). The deteriorated image quality resulting from SLC failure, which was mentioned before, has become a major obstacle for Landsat 7 data applications (Chen, Zhu, Vogelmann, Gao, & Jin, 2011b).
Toward the remote monitoring of armed conflicts
2023, PNAS NexusUnderstanding peace through the world news
2022, EPJ Data Science
- 1
Author is a foreign affairs analyst at the U.S. Department of State. He wrote this in his capacity as a graduate student at the University of Maryland. Any views expressed herein are solely his and do not necessarily represent those of the U.S. Department of State or the U.S. government.