Using object-based image analysis for automated displacement shelter identification and count
Worldwide displacement is on the rise, due to both natural disasters and complex emergencies, leading to a global displacement crisis with 65.3 million displaced people as of the end of 2015. The majority of those — 40.8 million — are Internally Displaced Persons (IDPs). The detection and count of displacement shelters is critical for effective camp management and the provision of necessary and timely aid and infrastructure. Automatic feature extraction based on remote sensing could save time, as detection and count based on visual inspection — while highly accurate — is highly time consuming.
This short study describes the feasibility of using very high resolution satellite image data and object-based image analysis (OBIA) to automatically detect and count displacement shelters related to complex emergencies. The study will look at IDPs camps only. IDP camps are often characterized by a highly complex, not uniform camp structure and limited access for international humanitarian organizations, as they are created in or in close proximity to areas of armed conflict. This complexity results in additional challenges for automated feature extraction and the transferability of methodologies to different camps and environments.
A successful deployment of OBIA for shelter detection and count could provide the following added value:
- Speed up the analysis time across multiple sites
- Monitor change over time at the same site
This research was conducted as part of Penn State University’s Remote Sensing Image Analysis and Applications course, and is a historical study not intended for operational use.
Data and Methods
An eCognition rule-set was developed to identify and count displacement shelters looking at a small IDP camp that is structured in a grid system. The goal of the analysis was not only to identify and count all structures within the camp, but also to separate shelters from administrative support structures, thus differentiating structures into two sub-classes. A sole structure count — while easier to implement — would severely limit the value of the research.
The rule-set was then tested at a different camp that is not organized in a grid system, to assess how transferable the rule-set is to different environments, and how much the rule-set will have to be adapted, which appears to be standard practice in this area of research: “[…] refugee camps adapt to different natural and political conditions and thus differ considerably in structure, materials and building density. Furthermore, variations due to different satellite sensors and data acquisition times have to be considered. Consequently, parameters have to be adapted to some degree to specific camps and datasets.” [source] This second step was intended to assess the usefulness of OBIA for shelter detection and count, as OBIA adds most value when scaled, instead of just aiding in the analysis of one specific site.
Pre-processing and mapping was conducted in ArcGIS 10.3.
Study area and data
The primary area of interest was the Sujjo/Sudjo IDP camp (36.623, 37.092) in northern Syria, close to the Turkish border. This camp is one of several planned camps in the area, and hosts IDPs who fled ongoing hostilities related to the Syrian armed conflict. The camp appears well organized and follows a grid system that is common for planned displacement camps. Such a system helps with camp organization and is characterized by the presence of space between shelters to provide access to vehicles, protect against fire, and promote hygiene. The camp is 12 hectares in size and consists of different types of shelters. The majority of the shelters consist of rectangle shaped structures, likely caravan trailers common in camps related to the Syrian crisis. The south-east corner of the camp consists of white tents, likely hexagonally shaped, covered with blue tarp. Administrative support structures are similar in color than shelters, however, are larger. The camp appears to be administered by Syrian or Turkish NGOs such as The Foundation for Human Rights and Freedoms and Humanitarian Relief (IHH). No official camp information was found in publically available sources.
A secondary area of interest, which was only used to assess the transferability of the OBIA rule-set to displacement camps in different environments and to imagery based on a different sensor, is the displacement camp outside the Sortoni UNAMID base (13.420, 24.342) in Darfur. This camp is an unplanned camp and emerged at the beginning of 2016 following a government offensive in the Jebel Marra region of Darfur, which lead to widespread and systematic destruction of civilian homes. Fleeing civilians sought shelter in the vicinity of the already existing UN base in Sortoni. The camp is not set up in an organized way. Shelters appear consistent across the camp, and consistent largely of square- or rectangular-shaped, brown structures built with local materials covered with white tarp provided by the UN Refugee Agency (UNHCR). The camp is supported by the United Nations and international relief organizations. 21,536 people are registered at the camp, however, the actual number of IDPs is likely higher, as registration as been suspended since February. The camp does not have any official boundaries.
Both images were received orthorectified. The Pleiades imagery was initially received as two images, one 0.47meter panchromatic image, and one 2m multispectral image. The data was pan-sharpened using the pan-sharpening function in ArcMap to obtain a new, multi-spectral image with the spatial resolution of the panchromatic image. The WorldView-2 image was received pan-sharpened. Since only a small portion of the imagery, i.e. the extents of the IDP camps, was relevant for the analysis, the images were clipped to a smaller area of interests in order to minimize processing time.
Rule-set development and transferability
A multi-resolution segmentation algorithm was used to perform the initial image object creation from the satellite data of the primary area of interest. A small scale parameter of 25 was used considering the small size of the features of interests, such as tents. For example, tents used in Zaatari (Jordan), which appear similar to the ones used in Sujjo, display dimensions of 6.6m x 3.8m. A low shape weight (0.2) was used to give more weight to the spectral information in the data. In a second segmentation step, a spectral difference algorithm was used. This second step was critical to merge smaller, neighboring objects with a similar spectral signature to create relatively compact image objects that are representative of the features of interest and suitable for analysis.
Following the segmentation, multiple classification steps were created. The classification process focused on spectra, geometry and context. The strong brightness values and consistent size of the caravans proved useful for classification. A major challenge was the classification of the white tents covered with blue tarp, which possess very low brightness values and a less compact shape. However, the blue tarp displayed relatively high values in the near infrared band compared to the optical bands. The resulting high maximum difference value, which represents the maximum difference between bands, allowed distinguishing these types of shelters from all other features.
The larger size of the administrative structures was used to distinguish them from shelters. Finally, shelters in a planned camp set up in a grid system means that shelters are placed in very close proximity with a consistent distance to each other. This fact was used to reclassify some misclassified outliers outside the camp, i.e. detected shelters that were not close to any other shelters were different structures and were thus unclassified.
The developed rule-set was applied to the satellite data of the Sortoni IDP camp in Darfur. In order to test transferability, only minor adjustments were made to account for the different environment, camp structure and sensor. For example, no blue tarp is used at Sortoni, so this classification step was completely removed, in addition to adjusting the brightness and size values. However, no new segmentation or classification steps were added.
Results and Discussion
A total of 1,174 structures were automatically extracted in Sujjo, of which 1,083 were shelters and 91 were administrative support structures. A few of the detected structures (13 in total) were outside the official camp extent, and were exclude from the final output. The number of shelters detected within the official camp extent represents a 12.1% increase from a previous analysis, based on visual inspection, conducted by UNOSAT. This is consistent with a visual inspection of the satellite images from the two different dates (June 2015 and February 2016), which reveals an increase of tents in the south-east section of the camp.
The rule-set provided generally good segmentation results, clearly separating many shelters and administrative structures from their environment. The resulting classification was able to correctly detect and classify a majority of the shelters, especially the caravans. However, there were some unsatisfactory segmentation results, most notably in the central administrative area of the camp, where some administrative structures where not separated from other features. This led to some challenges during the classification, and subsequently misclassification of some features.
A true accuracy assessment with field data is often not possible for research projects related to non-permissive environments. Independent visual inspection of the image data to create reference points is thus a common methodology for accuracy assessment for projects analyzing displacement camps. Accuracy reference points were created using the same imagery of the primary area of interest that was used for the rule-set development, and the previous UNOSAT analysis that independently identified shelters and administrative structures. The overall accuracy for the rule-set is 87%.
In Sortoni — the secondary area of interest used to test the transferability of the rule-set — a total of 3,670 structures were detected, of which 3,395 were classified as shelters and 275 classified as administrative support structures. This automated analysis was conducted with a minimal adaption of the rule-set. These results might be useful in the immediate onset of the displacement crisis when timely analysis is critical, as it only took a few minutes to acquire the analysis. However, more time consuming customization of the rule-set would be required to improve the results, questioning the usefulness of an automated approach. While the results provide a coarse snapshot of the humanitarian situation, the results likely overestimate the number of administrative support structures, and misclassify several buildings within the UN base as shelters.
Conclusion
The added value of automated feature extraction based on object-based image analysis stems from scaling the process to large areas, yielding time-saving results. Such time-saving benefits would be immensely useful in humanitarian relief operations, where the timely delivery of aid is one of the main objectives. Considering the results of the research presented here, the feasibility of an automated OBIA approach to displacement detection and count appears to be limited. While it would be possible to develop a rule-set to satisfactorily analyze a specific camp, the process is not directly transferable to other camps. For example, the same rule-set developed for the Sujjo IDP camp would need to be adapted for other camps in the same area, as they use shelters with different shape and materials. This non-homogeneity of shelters and dissimilarity of camp structures is a major reason why automated feature extraction using OBIA is currently rarely deployed by humanitarian organizations who conduct remote sensing analysis. An additional obstacle is the cost-prohibitive nature of software like eCognition for non-profit humanitarian or human rights organization. However, this encourages collaboration with academic experts in these field who have the necessary technical skills for effective rule-set development and access to the appropriate software.
It is important to note that scaling the automated process to multiple sites is not the only added value. A strong rule-set developed for a specific site could contribute to time efficiently monitoring the change in shelters over time at that site.
An obstacle — compared to remote sensing projects in other fields — is the limitation of available data. The use of LiDAR data and resulting surface models could potentially strongly improve analysis. As many displacement camps are located in non-permissive environments, LiDAR data in many cases would be limited to data collected from space-borne sensors, which results in limited availability of publically available data for research purposes.
Despite these limitations, the approach shows some promises, especially when narrowing down the project scope. Instead of aiming for the development of a general displacement shelter detection rule-set, processes could be developed for planned camps only that are characterized by a grid system. Such a focus could make the transfer of rule-sets more valuable. Additionally, it is important to note that more advanced segmentation and classification methods showed promising results in the past. An improvement to the rule-set development would be the use of relative brightness and size values, as opposed to absolute values used in the research presented here.
The development of the rule-set for this project was greatly aided by careful review of an existing image interpretation key and literature. A take away from the project is thus that one should not underestimate the time it takes for background research, which eventually made the workflow and rule-set development much easier. It also helped to distinguish the research from other projects, which are largely focused on shelter detection, but not on the detection of separate administrative structures.
