Damaged Building Detection with Improved Swin-Unet Model

Automatic detection of damaged buildings from satellite remote sensing data has become an urgent problem to rescue planners and military personnel. Unfortunately are in different dimensions shapes with roofs depending on the type material be painted. In this study, we present improved Swin-Unet approach that comprises three main operations. First, as a Unet-like pure Transformer is used for multitemporal image segmentation. Second, features extracted using hyperspectral classification algorithm. Finally, binary change map generated, evaluation results obtained. This article takes AIST building scene example, compared conventional approaches tested, overall accuracy, mean intersection over union, separated Kappa proposed method were by at least 23.36, 0.1725, 0.202, respectively. Furthermore, scenes, such Gaofen-2/Jilin-1 optical images imagery dataset (xBD), have also come same conclusion. Thus, it provides advantageous capabilities monitoring along coastal areas.