GROUNDBREAKING   

                                         EARTH OBSERVATION / GIS

NATURAL RESOURCE MANAGEMENT     

 

 

UNITY STATE - MONITORING MASSIVE HUMAN ABUSES

 

- The oil fields of Sudan / Southern Sudan -

 

 

Development of EO applications and monitoring system reveals massive land use changes in two periods 1999-2002 and 2014-2015

Erik Prins (2018) Landsat approaches to map agro-pastoral farming in the wetlands of southern Sudan, International Journal of Remote Sensing, 39:3, 854-878, DOI: 10.1080/01431161.2017.1392634

Background work and introduction here

 

During 2013-15 the 5a area was exposed to more humanitarian suffering. This was apparently watched by international organisations as a huge amount VHR data was available on Google Earth. This allowed a more comprehensive evaluation of the Landsat potential for mapping the footprint and changes in the agro-pastoral way of life.

Very High Resolution (VHR) satellite data was used to evaluate the prediction of farming activity, which was characterized  as in the Prins 2009 report (Table 1; Prins 2018). However, the lower bound of farming activity was set to 10% of a Landsat pixel, as the target was the settlement areas and close surroundings - and below this threshold interpretation of VHR data was considered to be linked with unacceptable uncertainties. 

 Table 1. Farming activity class definition and means to verification by VHR imagery.

The key Landsat data (table 2) was again chosen from early dry / harvest period where farming activity are concentrated around the homesteads - thus, if present, will show a footprint like monitoring the pioshere in rangeland management.

Table 2. Landsat data used in International Journal of Remote Sensing publication (Prins 2018) included two essential images from the Prins 2009 report. The images are requested in early dry season that is best for rangeland monitoring and in this case are when people are concentrated at homesteads for crop harvest.

The five Landsat images were calibrated to ToA and intercalibrated using invariant points, with a few exceptions reached a remarkably close inter-calibration that allow multi-temporal monitoring. Surface Reflectance (SR) that is a product that have under development over the past decade showed a similar result. SR is useful as it corrects for uneven atmosphere and can now be acquired pre-processed for free. The uneven atmosphere is typically affecting visible blue band, but Blue did not show channelling to calibrate for the time series analysis. This was found both in Prins (2018) (see table  3 in Prins 2018 - calibrated data can be found below) and in the imagery shown in the Stockholm 2013 presentation that were calibrated independently.

Table 3. Mean difference between spectral bands after inter-calibration (from Prins 2018).

The most successful EO application for predicting farming activity was the machine learning algorithm MaxEnt, which is traditionally used in ecology for species prediction (Phillips et al 2006). However, so far, the remote sensing community has only given MaxEnt limited attention. In this case it outperformed all other traditional remote sensing classifiers. MaxEnt showed outstanding predictability and across all years - apart from the NiR band. This could be expected as farming areas in this environment could include a range of factors that include bare areas, differences in soils, green vegetation, and moisture, which the Landsat NiR band is sensitive to.

 

Figure 1. AUC values [1]of individual Landsat bands produced by MaxEnt for each year and mean of all years including SD. Apart from NiR band it shows remarkably high consistency and thus high reliability for multi-temporal prediction and prediction of farming activity.

A comprehensive analysis of classifications algorithms as well as spectral bands and indices was preformed from both the 2014 and 2015 image. The predictability of farming activity from MaxEnt showed very strong across indices.

The strongest response collaborated findings from recent years remote sensing research in rangeland management (fig 2). SWIR-2 (Landsat band 7) and furthermore indices of the Normalized Difference Tillage Index (NDTI), Normalized Difference Infrared Index 7 (NDII7) as well as the Soil Adjusted Total Vegetation Index (SATVI) showed strongest – they are considered state of art indices for rangeland management. However, Blue band and SWIR-2 showed for both test years to independently produce among the highest overall mapping accuracy > 97% (see Prins 2018 paper and showed a high Kappa coefficient of approximately 0.85). MaxEnt using all bands and Tassel caps outperformed all other methods with overall mapping accuracy > 98.5% (very high Kappa 0.88-0.92). The ISO classifier produced overall mapping accuracy > 97.2% (high Kappa 0.81-0.86) that outperformed most categorical classifiers and thus confirms Tueller (1989) statement from that time, that unsupervised classifiers tend to work better than supervised for rang lands. However, it should be noted that the uncertainties with ISO classifier was linked with high commission errors. Or more specifically, it made the farming areas larger which indicates it grabbed farming activity below the 10% threshold (compare figure 3a and 6). As such, the EO application for monitoring the effect of grazing / pastoral utilisation has still much more to come for. Particular, in terms of quantifying lighter grazing pressure and untangling the behind laying factors that appears to be driven by vegetation cover and productivity, exposed and cattle trampling of soils as well as dryness.    

Figure 2. AUC values produced by MaxEnt for individual Landsat bands and indices showed very high explaining ability, particular for indices and bands that have found useful for Rangeland management.

When the farming prediction from MaxEnt was put together as a standardized difference image (2014–2015), depopulated or change areas could be identified at both settlement and state levels. As shown in Figure 3 (b), there has been a significant decrease in farming activity along the main roads as well as in the east central part of the study area. This includes the Ngop area that has been documented by UNOSAT (2015) as burnt and the larger Boaw area that was reported to be severely terrorized and destroyed in 2015 (Aljazeera 2015; HRW 2015; UNOSAT 2016). Interpreting a no comprehensive mapping from Boaw area (Figure 3 b) in Google Open Street Maps (2015) shows approximately 1500 housing structures and with average of 5 persons in a household (personal communication Nils Carstensen, Christian Aid), this can bring up realistic estimates of 10,000 people being driven off their land in the Boaw area in 2015. The Landsat change product shows a similar overall pattern as the UNOSAT (2016) VHR-based product of destroyed housing structures. However, the Landsat application covers an extortionate larger area, and still allows identification of specific destroyed settlements (Figure 5). In addition to the massive disturbance east of Ngop and Boaw, the Landsat change analysis showed large scale increased farming intensity in and around major towns of Bentiu, Koch, and Leer. According to ground reports (HRW 2015; UNHRC 2016), these areas have received high amount of displaced people as well as cattle in 2015. This means a higher natural resource preasure that is both difficult to observe or derive from VHR imagery. UNOSAT (2016) has used direct observation of cattles to assess peoples whereabouts, however, this only represents a snapshot of a dynamic situation. On the other hand, the Landsat application captures the footprint of the farming activity in terms intensity of the natural resource utilization.

Therefore, Landsat data can be a better choice for capturing piosphere or anthropogenic activity than the use of VHR imagery. This has analogy to evaluation of burnt areas where interpretation of Landsat data can be a better choice than VHR imagery (Sparks et al. 2015). Overall, these results suggest that Landsat data not only can be an effective supplement to VHR imagery but a more effective choice to produce regional overviews that allow evaluation down to the settlement level. However, it can also retrieve essential information that is not captured or cannot be interpreted by using VHR imagery. This includes natural resource utilisation and has direct relevance for sustainable development as well as  understanding and monitoring earth science systems .

Producing a similar MaxEnt standardized difference image for the 1999–2002 data (see Figure 3 (a)) uncovers a massive change in farming activity that closely collaborated Prins earlier work and ground reports (e.g., HRW 2003; de Guzman 2002 - see also hand drawn map herein; Christian Aid 2002) of massive human abuses that eventually depopulated the Nhialdiu area in February 2002. Inserted in Figure 3 (a) are also attacked villages and refugee campsites that were geo-located from the reports. Furthermore, in transparent, the extent of the 2002 attack interpreted from reports and hand drawn maps (de Guzman 2002; Christian Aid 2002) that were reported to be depopulated in 2002. This could again be shown by the MaxEnt change detection product (e.g., SD < −2.5) that inferred the decrease in farming activity for the entire area. Again, using Google Open Street maps un-comprehensive housing structure assessment from approximately 2013 (HOT 2016) that covered approximately two-thirds of the area and 14.450 housing structures. This suggest that numbers of approximately 100,000 people have been driven off their land for that area. Most people found refuge in the south (in green) that Christian Aid (2002) have assessed to receive approximately 50,000 people by late 2002. Inserted in Figure 3 (a) is a zoom up along the all-weather road from where specific villages have been reported attacked and show no longer farming activities in 2002. It should be noted that most villages were reported attacked around 1999 or before but stayed there untill 2000. 

Figure 3. Change detection (z-scores) of MaxEnt derived farming activity on a large scale (from Prins 2018), its detail and collaborate reports of massive displacement of people from two different periods and events (1999–2002 (a) and 2014–2014 (b)). Figure (b) Close up of Boaw that was compleatly destroyed in 2014 small green areas are havealy degraded areas that has not recovered. The Landsat version of Ngop can be compared with the Unosat approach (fig 24) - the Landsat showed more than Unosat approach that in-fact hide more settlement in SE corner - can be viewed in GE pro. Remake the blue arrow (in b), which refers to the location of figure 25. Further and not at least remarke the green displacement areas that collaborate that after comprehensive looting people fled up to Nhialdiu area as well as the footprint at Koch and Leer Area that also was a big hub for cattle looters going south.   

   

Figure 4 Clip of farming change in the Ngop village area 2014-2015. HR image used by UNOSAT 2015 counting 250 destroyed structures – see more  https://unosat-maps.web.cern.ch/SS/CE20131218SSD/Ngop_UNOSAT_20150518.pdf

Figure 5. A typical settlement area recorded by VHR imagery in late 2014 (a), cattle’s, most white, can be seen in upper right part of the image. The same area recorded 7 months later into the wet season. The settlement was destroyed and is rapidly being overgrown (b). The Landsat 2014–15 difference image (z-scores) of farming activity from MaxEnt shows a strong response to the change at settlement level. Photo credit (Google Earth).

 

Figure 6. Change detection of farming activity after harvest in 1999 and 2002. Based upon ISO classification of calibrated Landsat data (Prins 2018 and avaliable below) recorded the 27. Nov. 1999 and 3. Nov. 2002.

The ISO classification was performed with standard setting in Erdas Imagine - like the Prins 2009 report. The categorical classifier immediately grabs more of the homestead surroundings than using the supervised approach with a 10% minimum threshold. This means that it captures farming activity well below the 10% threshold set out in the VHR verification. In other words, the ISO classifier grabs the effect of farming activity that are difficult to account for in VHR images interpretations. This sensitivity is a plausible explanation for why this classifyer has for decades been considered as a stronger classifier than supervised classification to trace the effect of pastoral farming systems. As such, is a simple and transparant ML algorithm that can be strong to class land covers that appears complex to threshold on the ground or by supervised classifyers. This strainght are also known when it comes to untangle functional forest covers - see also  e.g., biodiversity session in ESA Living Planet Symposium 2023. 

Concluding remarks on the Landsat application

VHR imagery has a clear advantage over Landsat data of being able to check up the status of individual building structures at any time of the year if the cloud cover permits it. However, Landsat / Sentinel 2 data offer a huge potential for swift  assessment of humanitarian crises and derive information that can not be directly interpreted from of VHR imagery. A large part of this potential can be reached by understanding earth system processes which is far from being realised. It can not be emphasised enough - to unfold the potential - you need a good understanding of both human and natural ecology or the earth system. The result of using inter-calibration and continuous data has further prospect for monitoring heavy resource utilization that is essential for understanding sustainable resources utilization. This refers to ecosystem degradation processes that can have traumatic consequences if the resilience is broken (Holling 1973). This is another thematic area that have been far to little touch upon.

The spatial and multi-spectral properties of Landsat data were able to traced farming activity beyond what could be clearly interpreted from VHR imagery. For example, this can explain the commission errors (show too much) by the ISODATA classification that appeared to be rooted in farming activity below the 10% farming activity threshold used in this work.

The work unfolded the highly effective MaxEnt algorithm on the great potential of Landsat data. This out-perform all other tested commonly known EO classification approaches or algorithms. For example, this resulted in uncovering and detailed a human tragedy covering the area in 2014-15 - that appears to have missed the attention of media and  international community and thus to take appropriate measures. This is what EO is good at - to shed light on humanitarian crises - however, it is the responsibility of those big organisations that take on the EO application to show it.

If the result of the 2014-15 case is carefully interpreted together with ground reports, the spatial pattern (e.g. figure 23 b) tells a story which reveals what actually have happened and the extent of it. In other words, the scale of this result provides justice evidence and political power that interpretation of the Ngop case will never have. 

 

 

Supplements:

 

PRINS KEY-NOTE SPEAKER AT THE SWEDISH SPACE DAYS 2013: Global reporting interpretation of speach

 

  pdf  PRESENTAION OF THE 5A CASE AND ANALYSIS - INVITED TALK REMOTE SENSING DAYS STOCKHOLM 8 APRIL 2013

pdf  REPORT TO ECOS 2009: Satellite mapping of land cover and use in relation to Oil exploitation in concession 5A

 

Calibrated Landsat bands produced for and used in the Prins 2018 publication and analysis (files are in ASC formate)

 

 


[1] In general, an AUC of 0.5 suggests no discrimination (i.e., ability to predict farming activity), 0.7 to 0.8 is considered acceptable, 0.8 to 0.9 is considered excellent, and more than 0.9 is considered outstanding.

 

References:

Aljazeera. 2015. Accessed 26 September 2016. http://www.aljazeera.com/indepth/features/2015/ 05/south-sudan-man-catastrophe-150525120843494.html

Bastin, G. N., G. Pickup, V. H. Chewings, and G. Pearce. 1993. “Land Degradation Assessment in Central Australia Using a Grazing Gradient Method.” The Rangeland Journal 15 (2): 190–216. doi:10.1071/RJ9930190.

Booth, D.T., and P.T. Tueller. 2003. Rangeland monitoring using remote sensing. Arid Land Research and Management 17 (4): 455–467.

Christian Aid. 2002. Hiding between the Streams: The War on Civilians in the Oil Region of Southern Sudan. http://www.christian-aid.org.uk/indepth/0205suda/sudanrpt.pdf

de Guzman, D., 2002. Depopulating Sudan’s Oil Regions. European Coalition on Oil in Sudan (ECOS). http://www.ecosonline.org/back/pdf_reports/2002/depopulating%20sudans%20oil%20regions.pdf

Fagan, Matthew, and Ruth DeFries. "Measurement and monitoring of the world’s forests: A review and summary of remote sensing technical capability, 2009–2015." UMBC Geography and Environmental Systems Department Collection (2009).

Google Open Street Maps. 2015. “The Humanitarian Openstreetmap Team.” https://www.hotosm. Org

Holling, C. S. 1973. Resilience and stability of ecological systems. Annual Review of Ecology and Systematics 4:1-23

HOT. 2016. Project - South Sudan, Unity State, Bentiu. The Humanitarian OpenStreetMap Team. http:// tasks.hotosm.org/project/396

HRW. 2003. “Human Rights Watch (November 2003). Sudan, Oil and Human Rights. 754 p. Accessed 26 September 2016. http://www.hrw.org/reports/2003/sudan1103/9.htm#_ftn88#_ftn88

HRW. 2015. “’They Burned It All’ Destruction of Villages, Killings, and Sexual Violence in Unity State, South Sudan: Report.” 42 p. Accessed 26 September 2016. https://www.hrw.org/sites/default/ files/report_pdf/southsudan0715_web_0.pdf

James, C. D., J. Landsberg, and S. R. Morton. 1999. “Provision of Watering Points in the Australian Arid Zone: A Review of Effects on Biota.” Journal of Arid Environments 41 (1): 87–121. doi:10.1006/jare.1998.0467.

Lemenkova, Polina. "ISO Cluster classifier by ArcGIS for unsupervised classification of the Landsat TM image of Reykjavík." Bulletin of Natural Sciences Research 11.1 (2021): 29-37.

Maynard, C. L., R. L. Lawrence, G. A. Nielsen, and G. Decker. 2007a. “Ecological Site Descriptions and Remotely Sensed Imagery as a Tool for Rangeland Evaluation.” Canadian Journal of Remote Sensing 33 (2): 109–115. doi:10.5589/m07-014.

Maynard, C. L., R. L. Lawrence, G. A. Nielsen, and G. Decker. 2007b. “Modeling Vegetation Amount Using Bandwise Regression and Ecological Site Descriptions as an Alternative to Vegetation Indices.” GIScience & Remote Sensing 44 (1): 68–81. doi:10.2747/1548-1603.44.1.68.

Olofsson, P., G. M. Foody, M. Herold, S. V. Stehman, C. E. Woodcock, and M. A. Wulder. 2014. “Good Practices for Estimating Area and Assessing Accuracy of Land Change.” Remote Sensing of Environment 148: 42–57. doi: 10.1016/j.rse.2014.02.015.

Phillips, S. J., R. P. Anderson, and R. E. Schapire. 2006. “Maximum Entropy Modeling of Species Geographic Distributions.” Ecological Modelling 190 (3): 231–259. doi:10.1016/j. ecolmodel.2005.03.026.

Pontius, R. G. Jr, and M. Millones. 2011. “Death to Kappa: Birth of Quantity Disagreement and Allocation Disagreement for Accuracy Assessment.” International Journal of Remote Sensing 32 (15): 4407–4429. doi:10.1080/01431161.2011.552923

Erik Prins (2018) Landsat approaches to map agro-pastoral farming in the wetlands of southern Sudan, International Journal of Remote Sensing, 39:3, 854-878, DOI: 10.1080/01431161.2017.1392634

Samain, O., L. Kergoat, P. Hiernaux, F. Guichard, E. Mougin, F. Timouk, and F. Lavenu. 2008. “Analysis of the in Situ and MODIS Albedo Variability at Multiple Timescales in the Sahel.” Journal of Geophysical Research: Atmospheres 113: D14. doi:10.1029/2007JD009174

Stehman, S. V. 2004. “A Critical Evaluation of the Normalized Error Matrix in Map Accuracy Assessment.” Photogrammetric Engineering & Remote Sensing 70 (6): 743–751. doi:10.14358/ PERS.70.6.743

Tueller, P. T. 1989. “Technology for Rangeland Management.” Invited Synthesis Paper 42 (6): 442.

UNHRC. 2016. “UN Human Rights Council.” Assessment mission by the Office of the United ations High Commissioner for Human Rights to improve human rights, accountability, reconciliation and capacity in South Sudan: detailed findings, 10 March 2016, A/HRC/31/CRP.6. Accessed 26 April 2016. http://www.refworld.org/docid/56e2ee954.html

UNMISS 2015: UN Mission to South Sudan. 2015. “2015, Flash Human Rights Report on the Escalation of Fighting in Greater Upper Nile.” April/May, June 29. Accessed 26 April 2016. http://www.refworld.org/docid/5592995c4.html

Unosat. 2016. “UNOSAT LIVE Map for South Sudan Complex Emergency (CE20131218SSD).” Accessed 26 September 2016. https://unosat.maps.arcgis.com/apps/webappviewer/index.html id=cac78381476c43ebb98a7e642211809d

UNOSAT. 2015. “Rapid Damage Assessment of the Area of Rapid Damage Assessment of the Area of Ngop Ngop, Unity State, South Sudan.” Accessed 26 September 2016. http://unosat-maps. web.cern.ch/unosat-maps/SS/CE20131218SSD/Ngop_UNOSAT_20150518.pdf

Washington-Allen, R. A., N. E. West, R. Douglas Ramsey, and R. A. Efroymson. 2006. “A Protocol for Retrospective Remote Sensing–Based Ecological Monitoring of Rangelands.” Rangeland Ecology & Management 59 (1): 19–29. doi:10.2111/04-116R2.1.

 

 

 

privacy@prins engineering.com

About Us    |    References    |    Services    |    CV's    |    Contacts    |    Press     Copyright © PRINS, Eng. 2023