Automatic detection of collapsed buildings post-earthquake by deep learning and orthoimages

Earthquakes pose a serious threat to human life and safety through the structural destructions they cause. In this context, the rapid and accurate detection of collapsed buildings from high-resolution orthoimages obtained after earthquakes is of great importance for the effectiveness of disaster response processes. This study comparatively analyzes the performance of different variants (N, S, M, L, X) of the deep learning-based YOLOv12 model family in the task of collapsed building detection. The metrics from the training process revealed that all models underwent a stable learning process and did not exhibit overfitting tendencies. In particular, the YOLOv12-M model provided the most balanced results in terms of accuracy (AP: 0.940) and resource efficiency, while the L and X variants maintained similar levels of success but stood out with slightly higher recall rates in cases of increased scene complexity. Nonetheless, the similar accuracy levels of these three models offer flexible selection options depending on hardware and speed requirements based on the application scenario. Additionally, the study proposes a method capable of converting model outputs into geographic coordinates without adding extra processing load to the classical deep learning workflow. In this context, orthoimages were divided into 640×640 patches in both pixel and geographic coordinates, and the bounding box and center point coordinates of the detected objects were automatically obtained. Thus, the detected collapsed buildings have been made integrable with geographic-based decision support systems. The results show that with the proposed method, high-accuracy and spatially informed building detections can be achieved, supporting its applicability especially in areas where time and location sensitivity are critical, such as disaster management.