A Fast and Lightweight Method for Detecting Damaged Buildings After Earthquakes

Abstract

Earthquakes can cause significant damage to human-built structures. The rapid assessment of damaged buildings in disaster-stricken areas is crucial for making informed rescue decisions and managing emergency responses effectively. To address challenges such as the persistent difficulty in obtaining pre-earthquake images, the complex backgrounds of post-earthquake remote sensing images, and the low efficiency of existing detection models, this paper proposes an improved YOLOv8-OBB model for rapid damaged building detection using single-temporal post-earthquake remote sensing images, which enable direct feature extraction from imagery acquired at a specific post-event time point without relying on multi-temporal data fusion. Firstly, the quality of remote sensing images is enhanced using gamma correction and bilateral filtering techniques. Secondly, during the feature extraction stage, the PKI module and the CAA attention mechanism from PKINet are combined with the C2f module to enhance the model's ability to extract features of damaged buildings and capture long-range contextual information effectively, while maintaining a lightweight design. Finally, the Focal Loss strategy is introduced into the loss function to address the foreground-background class imbalance in the object detection task. This approach enabled the model to focus more on hard-to-detect samples, thereby improving its recall rate. Experimental results demonstrated that the proposed model exhibited comprehensive advantages in damaged building detection tasks, achieving significant improvements in precision, recall, mean Average Precision (mAP), inference speed, and lightweight architecture compared to both baseline models and state-of-the-art algorithms. This advancement provided a superior technical solution for real-time analysis of post-disaster remote sensing imagery.