Deep-Learning-Based Assessment of Building Damage and Road Debris Following the Kahramanmaraş Earthquakes Using High-Resolution Remote Sensing Imagery

Earthquakes can leave large urban areas with collapsed buildings and debris on roads, making rapid damage mapping essential for emergency response and recovery planning. This study compares four semantic segmentation models, including ResUNet++, UNet++, Attention U-Net, and DeepLabV3+, for the simultaneous extraction of collapsed buildings and road debris from high-resolution satellite imagery acquired after the 2023 Kahramanmaraş earthquakes in Türkiye. The experiments use three-band RGB imagery from the MAXAR Open Data Program together with pixel-level annotations for the two target classes. All models were trained and evaluated under the same experimental settings to ensure a fair comparison. Segmentation performance was assessed using Overall Accuracy, Recall, and Intersection over Union (IoU). Among the evaluated models, Attention U-Net achieved the highest performance, with an Overall Accuracy of 93.43%, a Recall of 82.76%, and an IoU of 89.13%. ResUNet++, UNet++, and DeepLabV3+ produced lower scores under the same evaluation protocol. The comparison shows that attention-based feature aggregation improves the delineation of earthquake-related objects in high-resolution RGB satellite imagery, particularly where collapsed structures and road debris exhibit irregular shapes and complex spatial patterns. These results provide an evidence-based comparison of commonly used semantic segmentation models for post-earthquake damage mapping. They may assist researchers and emergency management agencies in selecting suitable models for operational remote sensing applications.