Building Change Detection Based On Deep Feature Fusion Of Siam Network

As the level of township continues to accelerate,the urban landscape has undergone great changes,among which building changes are the most abundant.To carry out urban building change detection based on remote sensing technology to avoid illegal land use and illegal expansion is an important direction in the field of natural resources supervision at present.In order to address the problems of loss of depth features,the need to improve the accuracy of change detection and the difficulty of detecting building height changes in remote sensing image building change detection using deep learning,the research on building change detection is carried out in two dimensions,horizontal and vertical,respectively,and the algorithm of building change detection based on deep learning in satellite remote sensing is improved.The research in this paper is as follows:(1)The design and validation of a binary classification model based on siamese network depth feature fusion for high-resolution remote sensing images is implemented in the horizontal dimension.The siamese network is built based on a full convolutional neural network with fused residual structure,and the feature alignment module is used to modify the depth features to make the detection results clear and the feature pyramid module is used to the multi-scale building change detection is achieved by using a feature pyramid module.The experimental results show that the proposed network improves accuracy by 6.74%,recall by 6.08% and F1 score by 6.91% in the study area compared to the baseline network.Compared to the DTCDSCN network the accuracy rate improved by 2.42%,the recall rate improved by 5.8% and the F1 score improved by 6.1%.On the LEVIR-CD open source dataset the accuracy rate improved by 6.98%,the recall rate by 2.84% and the F1 score by4.31% compared to the CDNet network.(2)The design and validation of a multi-classification model for high-resolution remote sensing images based on siamese network depth feature fusion was achieved in the horizontal dimension.Firstly,a multiclassification building change detection dataset was produced using ENVI and Arc GIS Pro software,and a multiclassification building change detection model based on siamese networks was proposed to fuse depth features based on U-Net networks with added residual structure,and a null convolution module was used to achieve multiscale building change detection,and an attention mechanism was used to achieve the recognition of small targets in complex scenes.Compared with the change detection network based on U-Net and Res-UNet,the accuracy rate is improved by 9.4%,the recall rate by 10.6% and the F1 score by 9.8%.(3)Faster R-CNN and structural similarity based building height change detection and validation of high altitude lookout images were implemented in the vertical dimension.Firstly,a building dataset was produced using Labelimg software,and a building height detection method based on aerial lookout images was proposed.SSIM-based building change detection method was used to segment the two period images into N×N image subblocks,and then adaptive thresholding was used to roughly extract the change region,and experiments showed that the best building change detection accuracy was achieved when the segmentation scale was 7.Finally,the improved multi-scale Faster R-CNN network is used for target detection of buildings.Experiments show that compared with the difference method,the false detection rate and missed detection rate decrease by 12.7% and 51.8%respectively,and the correct rate,completeness rate and detection quality increase by 22.4%,51.9% and 12.0% respectively.