Research On The Algorithm Of Building Damage Detection And Change Detection Based On Remote Sensing Images

With the rapid development of remote sensing technology,the spatial resolution of remote sensing images is getting higher and higher,and the high-resolution remote sensing images have more plentiful ground information.The detection of building damage based on high-resolution remote sensing images pre-and post-disasters is of great significance in assisting post-disaster rescue and post-disaster response,while the detection of building changes based on dual-time phase remote sensing images plays a key role in urban and rural planning and land resource planning.With the constant advancement of deep learning algorithms,deep learning methods have achieved certain results in both building damage detection and change detection.For the task of building damage detection,background of high-resolution remote sensing images is complex,the shape,size,and distribution of buildings are more diverse,and the characteristics of adjacent damage categories are similar,resulting in the inaccurate positioning and classification of damaged buildings.In this paper,a two-stage building damage detection network is designed.The first stage is a building localization network,which adopts the Asymmetric Convolution Block and Multi Scale Jump Connection to improve the extraction capability of complex buildings and uses only pre-disaster images to complete the building localization work.The second stage is the building damage degree classification network,based on the first stage network to build a Siamese Network,embedded attention module to improve the correlation of buildings pre-and post-disasters,and remote sensing images pre-and post-disasters as input to complete the four classifications of building damage degree.For the building change detection task,the detection network is lacking in the acquisition of global information,and the existing building change detection network has a high number of parameters.In this paper,we design the lightweight EfficientUNet++ building change detection network to accomplish building change detection using semantic segmentation.The remote sensing images of dual time phases are first stitched together and then fed into the network.Self-attention is embedded at the bottom of the network to improve the network’s access to global information.A lightweight feature extraction network and Deep Separable Convolution are used to ensure the accuracy of building change detection while reducing the number of parameters in the network.