Damaged Building Detection with Improved Swin-Unet Model

Abstract

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