Uncertainty-Aware 2d/3d Change Detection for Natural Disaster Response

In the scope of crisis management support, CNES develops a change detection tool that exploits the hybridization of 2D and 3D satellite data in order to address the growing number of climate-related hazards and natural disasters around the globe: A 2D change detection pipeline, based on recent deep neural networks, detects semantic changes between very high-resolution satellite images. Additionally, a complementary 3D change detection pipeline confirms and completes the detected changes through data fusion schemes. We present the integration of uncertainty estimation and ambiguity concepts for both 2D semantic segmentation and 3D reconstruction in order to introduce a qualification of multi-modal change detection results and improve their reliability. The paper includes an evaluation of our method’s capacities to support natural disaster response in the context of the recent earthquakes in Turkey and wildfires in Chile, reviewed by experts who work on mapping activities for the International Charter Space and Major Disasters.