Detecting collapsed buildings caused by earthquake from remote sensing image based on deep learning

After an earthquake, rapid detection of collapsed building distribution from remote sensing images is crucial for government agencies, as it facilitates effective rescue operations and minimizes casualties. Traditional methods of damage assessment methods, such as on-site surveys and manual interpretation of remote sensing data, are often hindered by their labor-intensive and time-consuming nature, thus necessitating a more efficient approach. In this study, we address this challenge by employing a deep learning-based methodology for detecting collapsed buildings from satellite imagery. Specifically, we developed a Deep Learning model based on the U-net architecture, chosen for its high flexibility and performance. Utilizing a custom weighted loss function, our model achieved an Average Precision of 0.871 and an Average Recall of 0.893 on the test set. Application of the model to the Turkey earthquake case demonstrated rapid and accurate segmentation of most buildings. This study suggests that integrating U-net-based deep learning with satellite images can provide the precise distribution of collapsed buildings needed for earthquake emergency management.