A Study On Building Extraction And Change Detection From Remote Sensing Images Based On Semantic Segmentation

The acceleration of urbanization has brought about problems of illegal land occupation and illegal construction,which seriously affects urban construction and social sustainable development.Compared with manual inspection,remote sensing images contain rich ground object information,which has the advantages of fast information acquisition,short update cycle,dynamic monitoring and so on.Therefore,the paper studies the building extraction and change detection algorithm based on high-resolution remote sensing image combined with image processing and deep learning theory,so as to realize the automatic detection of building extraction and change in remote sensing image,and ensure the real-time tracking of urban construction process and change.The main research contents are as follows:(1)The semantic segmentation data set of high-resolution remote sensing image is established.According to the task requirements of building extraction and change detection from remote sensing images in this paper,a single static building detection semantic segmentation data set with the 2017 remote sensing image of Foshan City,Guangdong Province as the original image and a building change detection data set with new buildings as the semantic label are established respectively.At the same time,the establishment process of semantic segmentation data set is described in detail,including data set acquisition,remote sensing image preprocessing and data set production.(2)A building extraction algorithm based on SKDD-Deep Labv3+ is proposed.Aiming at the problems of missing building boundary information and missing detection of small target objects caused by continuous down sampling operation in the coding stage,channel attention,channel space attention and convolution kernel attention mechanisms are introduced into each residual module of feature extraction network Res Net101.SE-Res Net101,CBAM-Res Net101 and SK-Res Net101 networks are designed successively.The experiments show that the SK-Res Net101 can better improve the processing performance of feature extraction network in detail;Designing dense connected spatial pyramid structure to provide more effective context coding information;The DUpsampling method is used instead of bilinear interpolation to improve the prediction ability of the model to recover the feature map pixel by pixel.Experiments show that the algorithm effectively improves the segmentation performance of the model.92.75% MIo U was obtained on the WHU building dataset test set,and 89.23% MIo U was obtained on the building detection dataset of remote sensing images in Foshan City,Guangdong Province in 2017.(3)An end-to-end building change detection method based on Swin-UNet is proposed.Taking UNet network as the basic framework,a layered Swin Transformer with shift window is designed to replace the traditional CNN self coding module of UNet,so as to realize the complementary advantages of the network in which the Transformer pays attention to processing global information in the coding stage and CNN effectively extracts the underlying features and visual structure in the decoding stage.The experiments of single image and large-scale remote sensing image detection are designed.The experimental results show that this method has better change detection performance.84.72% MIo U is obtained on the LEVIR-CD public data set,which is better than other advanced algorithms.