Deep learning for automatic post-disaster debris identification for precise damage assessments using UAV footage

Abstract

With increased frequency and intensity of extreme climate events, unprecedented volumes of debris are created. Disaster debris can often be hazardous, and it obstructs relief activities by blocking the roads and preventing access to disaster sites. This highlights the importance of timely debris identification and removal efforts for effective relief and preliminary damage assessments. This work aims to automatically extract post-disaster debris from UAV footage using instance segmentation with YOLOv8-seg model. Automatic detection of debris, its type and geographical distribution allows efficient allocation of resources, prioritization of relief efforts and significant reduction in the time taken for disaster recovery. We use UAV images of Hurricane IAN, specifically of Julies Island along the coast of Florida. We trained and compared YOLOv8n (nano), YOLOv8m (medium), and YOLOv8x (extra-large) model architectures, to select the suitable model for post-disaster debris detection. Since debris clearance efforts typically depend on debris type, we trained and built specialized models for vegetation and non-vegetation debris separately. The YOLOv8x model exhibited the highest accuracy—83% accuracy for vegetation debris and 85% for non-vegetation debris, with corresponding mAP values of 62.2 and 66.1, respectively. The model detected non-vegetative debris as small as 0.13 square meters. Furthermore, we used YOLOv8 model to detect and track damaged, hazardous and non-hazardous assets on the street from UAV videos. We developed an algorithm to automatically produce georeferenced results from UAV images, enhancing the model's usability in real-world applications. The developed model automatically outputs precise location, size and area of debris, aiding post-disaster planning.