| Buildings provide basic functional space for urban residents to live,work,relax,and other activities.Rapidly extracting building information from high-resolution remote sensing image data is of great importance for urban planning,surface dynamic monitoring,and land use management.significance.However,considering the complex composition of remote sensing images,only relying on human interpretation is not only inefficient,but also has great instability.At present,the relevant building extraction and change detection algorithms have been able to realize the automatic segmentation of buildings and effectively improve the accuracy of building segmentation.However,due to the high distinguishability of features such as color and texture of buildings,certain spatial arrangement and distribution,different scales,complex edge shapes,etc.,and the characteristics of temporal correlation between the front and rear phase images in the change detection task,the building’s Research on automatic extraction and change detection algorithms is still a challenging task.Therefore,this dissertation focuses on building extraction and change detection models in remote sensing images.The main work includes:(1)Aiming at the problems of difficult location of building subjects in high-resolution remote sensing images,ineffective use of building spatial context information,and difficulty in fully extracting multi-scale buildings,a building extraction model based on a two-way supervision network is proposed—— BDS-UNet.First of all,the skip connection part of its U-Net model introduces a coordinated attention gating module,which aggregates feature information along two different spatial directions and complements each other to generate an attention heat map,reasonably integrates building features and realizes building Precise positioning of the subject.Secondly,a continuous spatial pyramid module is added to the bridging part of the U-Net model to group the small-scale hollow convolutions and superimpose the input features step by step within the group to further capture the spatial context information existing between different buildings.Finally,a bidirectional supervision structure is proposed,which divides the decoding stage into top-down and bottom-up two parts to propagate and fuse the prediction results of adjacent decoders,effectively improving the multi-scale building extraction ability of the model.Experimental results on the publicly available Massachusetts dataset and WHU dataset demonstrate the excellent performance of the proposed model in building extraction.(2)Aiming at the problems that it is difficult to fully capture the timing characteristics of the front and back temporal images,and the edge details and spatial context information of the changed area are easily ignored,a building change detection model based on change perception and global enhancement—CAGE-Siam-UNet is proposed..First,a change perception module is added to the feature fusion stage of the twin encoders of the Siam-UNet model,and the front and rear phase features are activated from the three branches of spatiotemporal activation,channel activation,and perceptual activation,thereby highlighting the differences in the changes of buildings in different time series.Secondly,a global enhanced decoder structure is proposed,through local refinement operation,spatial context feature fusion and bidirectional deep supervision,while enriching the edge details and spatial context information of the changing region,the multi-scale changing region extraction ability of the model is optimized.The experimental results on the public LEVIR dataset and the WHU change detection dataset demonstrate the excellent performance of the proposed model in building change detection. |
