A Novel Two-Stage Network for Building Localization and Damage Level Assessment

Automatically assessing the building damage level in natural disasters is critical for humanitarian assistance and emergency response. In this paper, we propose a novel two-stage method to assess the building damage level. Firstly, a deep neural network (DNN) is designed to segment the buildings from the satellite imagery. Then the output layer of the DNN is appropriately modified to extract the deep features of the predisaster and post-disaster imageries. Finally, a cross fusion module is proposed to fuse the corresponding deep features for assessing the building damage level. Compared with the directly adding, subtracting or concatenating pre-disaster and post-disaster features, the proposed cross fusion module can effectively reduce the impact of pre-disaster and post-disaster imageries misalignment and further improves the accuracy of the assessment. Our method surpasses the previous two-stage and end-to-end methods, and achieves state-of-the-art performance.