Research On Building Extraction Method Of Remote Sensing Image Using U-configuration Neural Network With Attention Mechanis

The fast and efficient extraction of buildings from remote sensing images is of great significance for map updating,urban change monitoring,urban planning and 3D modelling.The complex and diverse features of features in high-resolution remote sensing images make it difficult for traditional algorithms to extract buildings in a batch in an adaptive manner,therefore,the design of automatic building extraction algorithms has become a hot research topic of widespread interest.The superior performance of deep learning semantic segmentation provides a new idea for building extraction from remote sensing images,however,the large variation of features such as building styles,spectra and shadows leads to major problems such as inaccurate edges,tiny buildings missing and voids during building extraction.To address the above problems,the SERUNet and FSA feature self-attentive building extraction algorithms are proposed from three aspects of attention,multi-scale and feature fusion,combined with the U-configuration,where the SERUNet algorithm is used to improve the parallel multi-path network and the FSA algorithm is used to construct the FSAU-Net network,which is better than UNet,UNet++,PSPNet,Seg Net,FCN-8S,Res Net101 and other networks to effectively improve the building extraction accuracy.The main work is as follows:(1)A building feature extraction SER-UNet algorithm with channel attention mechanism is proposed to achieve the fusion of shallow features with deep features in the feature extraction phase of the network.The method uses a channel-attentive residual structure to extract shallow building features in the encoder stage and a channel-attentive residual structure to extract deep building features in the decoder stage,and then uses a jump connection to fuse the corresponding layers of deep and shallow features.This method is cascaded into a parallel multi-path network,so that the network outputs a multi-scale feature map,and the network focuses on the integrity of tiny buildings and building edges while effectively reducing voids,effectively improving the building extraction accuracy.(2)In order to verify the effectiveness of the proposed SER-UNet method and its enhancement of the parallel multi-path network,two experiments were conducted on the SERUNet method and the network respectively.Method Experiment 1: Experiment on the scale size of the method itself to find an optimal scale under multi-path and normal networks to maximise network performance;Method Experiment 2: Experiment on the generalisability of SER-UNet by cascading the method into different networks respectively to check whether the method is generalisable to different networks for performance improvement.Network Experiment 1:Building extraction visualisation and accuracy analysis on three publicly available datasets,WHU,Inria and Aerial Image,respectively,and comparison with comparison networks;Network Experiment 2: Experiment on network complexity and comparison with commonly used networks.The experiments show that the SER-UNet method is highly generalisable and that the building extraction accuracy of the parallel multi-path network combined with the method is greatly improved.(3)A self-attentive FSA algorithm is proposed to focus on building features,and the FSA is used as the core module to build a building feature extraction network,FSAU-Net,which couples the FSA algorithm in the coding stage.Spatial Attention,SA)method is used to focus on the spatial location of building features to improve the building edge accuracy and reduce voids.Features with high accuracy are fused with features with spatial location in the decoding stage to effectively improve the building extraction accuracy without increasing the number of model parameters.(4)The superiority of the method FSAU-Net is verified on WHU and Inria datasets with a resolution of 0.3m.Experiments show that the Io U reaches 91.73 and 80.73%,and the accuracy(Precesion)reaches 93.60% and 90.71%,respectively,which is a significant improvement over the Io U and accuracy of the comparison network.Meanwhile,to verify the superiority of the proposed FSA method,the FSA module and the SE and ECA modules with channel attention were added on the basis of UNet and Res Net101,in which,UNet+FSA improved the Io U by 3.15%,2.72% and 1.77% compared with UNet,UNet+SE and UNet+ECA,respectively,and Res Net101+FSA improves 2.06%,1.17%,and 0.9% over Res Net101,Res Net101+SE,and Res Net101+ECA,respectively,demonstrating the superiority of our proposed FSA module,which has a greater improvement on the model capability.