Research On Surface Coverage Monitoring Algorithm Based On Self-attention Network

Building and road extraction is an important task in remote sensing image land cover monitoring.Accurate segmentation of buildings and roads provides a great help for the reasonable layout of the city and is of great significance to the realization of urban planning.Most of the existing building and road extraction models use deep learning semantic segmentation methods,but these methods have some shortcomings.On the one hand,the feature fusion process directly stitches the feature map in the channel dimension,and the feature information between the hidden layers(channel dimension)is not paid enough attention,which leads to the neglect of the category of context pixels in the pixel classification,resulting in the problems of building large area misjudgment and road extraction disconnection.In order to solve this problem,this paper proposes a Non-Local Feature Search Network,and strengthens the exploration of hidden layer feature information through the attention feature transfer module,which effectively reduces the large-area misclassification of buildings and road extraction disconnections.On the other hand,there are limitations in the receptive field in high-resolution remote sensing images,which results in the lack of long-distance scene understanding capabilities during pixel classification,and the feature map is compressed during downsampling,resulting in loss of detailed information.In order to solve these problems,this paper proposes a Hybrid Multi-Resolution and Tansformer semantic extraction network.Through transformer semantic extraction,the global receptive field of the feature map can be obtained,which helps to obtain the long-distance scene dependence of the segmentation target,and uses the multi-resolution semantic extraction branch to overcome feature maps cause the loss of details in the process of downsampling.Finally,after a large number of comparative experiments on two remote sensing image semantic segmentation datasets,the experimental results achieved satisfactory segmentation accuracy,so the effectiveness of the proposed algorithm is verified.