Supporting Conflict Damage Assessment with Object-Based Image Change Analysis

Timely and detailed information on violent conflicts is essential to assess their impact on civilians and to document human rights issues, such as large-scale displacement. Remote sensing provides unique capabilities to overcome limitations of ground-based conflict documentation, especially in remote and insecure areas. Because the manual analysis of remote sensing data is time-consuming and labor intensive, this thesis investigates the use of geographic object-based image analysis (GEOBIA) for automatic damage assessment to support conflict monitoring. This includes the development of a suitable destruction detection method, but also concepts for the deployment and integration of GEOBIA in practical workflows of conflict analysts. We developed a GEOBIA method for the automatic and selective detection of destructed dwelling structures in bi-temporal very high-resolution images of two study areas in Darfur, Sudan. The method takes advantage of object-specific attributes, object hierarchies and relational features to increase the accuracy and the robustness towards differing image properties. It detected target objects with good accuracies in both study areas although three different sensors were involved. To lower the barriers to practical adoption of new GEOBIA methods in conflict analysis, we developed a concept combining free and open-source software (FOSS), parameterized Docker containers, and a graphical user interface. This approach allows to develop image analysis workflows without licensing costs that are transparent and reproducible, but also easy to use and to adapt by conflict analysts. As a result, it can advance collaborative development of GEOBIA methods, but also their practical adoption in conflict analysis. The thesis also presents a concept for an effective integration of results of automatic image analysis into workflows of conflict damage assessment. Because of the specific accuracy requirements in this domain and the uncertainty associated with automatic methods, results cannot be used as unmediated evidence. Instead, we developed a concept that uses GEOBIA to automatically prioritize subsets of a scene based on the detected degree of destruction and to guide human analysts to the most important locations. We implemented this concept as a prototypical web-application for collaborative conflict damage assessment. We conducted a user experiment to investigate and compare the performance of human analysts with and without assistance by automatic prioritization. In this experiment, the participants who were guided by the results of the GEOBIA method detected 70.7%more target objects than participants without guidance.