DS-Net: A dedicated approach for collapsed building detection from post-event airborne point clouds

Collapsed buildings should be detected immediately after earthquakes for humanitarian assistance and post-disaster recovery. Automatic collapsed building detection using deep learning has recently become increasingly popular because of its superior ability to obtain discriminative feature representations. Among various types of data, airborne 3D point clouds are especially useful for detecting collapsed buildings as they precisely record the height information of buildings. However, existing methods are based on the universal point cloud analysis technology that does not explicitly consider the nature of building damage. In this study, we propose Damage-Sensitive Network (DS-Net), a dedicated approach for collapsed building detection. The core of DS-Net is Laplacian Unit (LU), a simple yet effective module for 3D point clouds designed to enhance the feature representation of the damaged part to facilitate collapsed building detection. We perform extensive experiments and demonstrate that DS-Net achieves superior performance compared with existing methods. In particular, a detailed comparison of DS-Net with PointNet++, the standard network on which DS-Net’s design is based, found that DS-Net provides an 8.3% gain in precision, 3.0% gain in recall, and 6.4% gain in IoU over PointNet++ in detecting collapsed buildings. Moreover, it is verified that the detection performance can be further enhanced with increased computational resources. Qualitative analyses reveal that DS-Net excels at detecting damage manifested as roof deformations, debris, and inclinations. In addition, DS-Net produces smoother predictions with sharper boundaries compared to the baseline due to the adaptive nature of LUs. Furthermore, a visual explanation analysis based on Grad-CAM is performed to analyze how DS-Net understands building damage. The result suggests that DS-Net can accurately locate varieties of building damage. •We propose Damage-Sensitive Network (DS-Net) to tackle collapsed building detection.•The core of DS-Net is Laplacian Unit (LU), which adaptively identifies and enhances the damaged part.•Quantitatively, DS-Net significantly outperforms several mainstream deep learning-based methods for 3D point cloud analysis.•Qualitatively, DS-Net successfully detects damage-like patterns such as roof deformations, debris, and inclinations.•We provide a visual explanation analysis and show that DS-Net precisely locates various types of building damage.