Building Change Detection In High Resolution Remote Sensing Images Based On Spatiotemporal Attention Depth Perception

Building,one of the most basic elements on the Earth's surface,is the hottest research objective at home and abroad.It is of great significance to master the update of buildings in the fields of the urban monitoring,post disaster damage assessment and GIS map updating,etc.Nowadays,with the improvement of spatial resolution,high-resolution remote sensing images can provide abundant spatial texture information,having become a central data source for building change detection.However,the traditional change detection methods for building are inefficient and limited by the expression ability of artificial design features,so they are not suitable for identifying complex details,semantic and spatiotemporal features in highresolution remote sensing images.Therefore,how to carry out automatic and accurate building change detection based on high-resolution remote sensing images is an important problem to be solved urgently.The deep learning technology has the great potential in the image processing owing to its advantages in extracting depth features efficiently and accurately,so that it also has been widely used in building change detection in remote sensing images.From the perspective of semantic segmentation,this thesis carried out the research of building change detection in high resolution remote sensing image by deep learning.However,this method is easy to split the spatiotemporal correlation of dual temporal image and still unable to recognize the "pseudo change" in the image better,decresing the accuracy of the building change detection.In view of the above problems,the main work and conclusions of this paper are summarized as follows.(1)The spatial-temporal correlation extracted by U-Net model is not enough or to be ignored,making it difficult to identify the building areas changed in the dual phase image with "pseudo change".Aiming at the problem,a multi-feature fusion mechanism(MF)was constructed by combining the advantages of early fusion and later fusion.Furthermore,with the help of residual unit,the building change detection model MF?Res Unet based on the fusion of space-time characteristic was constructed.It could improve the representativeness,diversity and completeness of the spatio-temporal correlation characteristics of buildings,enhance the ability to identify the "pseudo change" of buildings,and improve the accuracy of building change detection.(2)For the MF?Res Unet model,it is easy to produce redundancy in the fusion and reconstruction of spatio-temporal correlation features of heterogeneous buildings,which leads to the poor training effect of the model and the decline of change detection accuracy.Aiming at the problem,the shuffle attention was introduced to construct spatiotemporal attention perception module(STAP),and the deep supervision module was used to design the monitoring module of the loss of depth preference(PDS).Finally,two modules were embedded in MF?Res Unet,designing a new method of building change detection based on spatial and temporal attention depth perception.While enhancing the effective fusion of building spatio-temporal association features,it could alleviate the influence of gradient non circulation and insufficient supervision information on the interaction of spatio-temporal association features in the end-to-end integrated network,improve the training effect of the model,and further improve the recognition ability of the model for building change areas.It is shown that,(1)In building change detection methods based on U-Net model,compared with the temporal and spatial correlation feature extraction mode of early fusion and later fusion,MF mechanism proposed in this paper is able to identify the "pseudo change" of buildings more correctly.(2)Compared with the traditional feature interaction model and supervision mode in MF?Res Unet,the STAP module and PDS module are able to further optimize the effect of the model to identify the change area of buildings.(3)By improving the model of deep learning and model training based on U-Net,a new method of building change detection based on spatial and temporal attention depth perception is obtained.Compared with the improved method,it has better detection results and higher accuracy in the detection of building changes in high resolution remote sensing images.The method of building change detection based on the depth of space-time attention has good effect on building change detection,and it is expected to better serve the urban monitoring,post disaster damage assessment,maps update and other work.The research work of this paper has certain reference value for improving the accuracy of building change detection of high-resolution remote sensing images.