SurfaceNet: leveraging aerial LIDAR point clouds for post-earthquake building damage assessment

Rapid mapping of post-earthquake damage is important for effectively staging rescue efforts, given that 90% of victims do not survive more than three days when trapped. Consequently, computational approaches to damage mapping have been explored to improve speed. Automatic building damage detection methods can produce results faster than human-based assessments, which has made them increasingly popular in both research and real-world applications. Applications of aerial LIDAR data in particular show high potential in this regard, because aerial LIDAR data can be obtained relatively quickly and contains valuable geometric information, which is inherent to patterns of collapsed buildings. Deep learning approaches have been recently explored for assessing the damage to buildings. However, most general point cloud deep networks have been designed using autonomous driving datasets and multi-view single-object datasets in mind, which negatively impacts their performance on assessing the damage to buildings from aerial LIDAR data. Consequently, we aimed to design an architecture embedding prior domain knowledge of aerial LIDAR for building damage assessment. We propose SurfaceConv, a novel local aggregator module, which enables consideration of long-ranged intra-surface dependencies and thus, a holistic consideration of distinct surfaces. Using the SurfaceConv module, we created SurfaceNet, a deep neural network (DNN) specifically tailored for detecting damage to buildings using aerial LIDAR data. We achieved results in the previously explored patch segmentation task that were comparable to existing models. We introduced the classification and general segmentation task and found that our model empirically outperformed all existing models. These existing models that we benchmarked against included both generalized DNNs trained from scratch, generalized pretrained DNNs, and building damage assessment specific DNNs.