Building Damage Assessment Based on Siamese Hierarchical Transformer Framework

The rapid and accurate damage assessment of buildings plays a critical role in disaster response. Based on pairs pre- post-disaster remote sensing images, effective building level can be conducted. However, most existing methods are based Convolutional Neural Network, which has limited ability to learn the global context. An attention mechanism helps ameliorate this problem. Hierarchical Transformer powerful potential field with strong modeling capability. In paper, we propose novel two-stage framework called SDAFormer, embeds symmetric hierarchical into siamese U-Net-like network. first stage, pre-disaster image is fed segmentation network for localization. second two-branch classification established weights shared from stage. Then, images delivered separately assessment. Moreover, spatial fusion module designed improve feature representation capability by pixel-level correlation, establishes information Swin blocks. proposed achieves significant improvement large-scale dataset—xBD.