Improving UNet+++ Network For High Resolution Remote Sensing Image Change Detection

High-resolution remote sensing images have more colorful geometric image data information such as texture features and adjacency relationships,which can provide a reliable source of information for intelligent interpretation of remote sensing images,and deepen their application in Change Detection(CD).Existing CD algorithms face huge challenges:First,the network relies on layer-by-layer convolution for feature extraction,lacks the mining of important changing features,and the upsampling process lacks effective connections,resulting in limited changing features learned by the network;second,the network is affected by Due to noise interference,it is difficult to fully exploit the difference information between images while obtaining image features;third,the problem of difficult and easy sample classification in the training process will affect the network accuracy.Therefore,this paper deeply analyzes the basis of the current high-resolution remote sensing image change detection technology In the above,combined with deep learning technology,the CD task is studied.The specific research contents are as follows:(1)By adding the Coordination Attention(CA)mechanism to the segmentation network UNet+++,the full-scale semantics from the encoder is fully learned and the problem of insufficient information acquisition in the image feature extraction process is solved.Comparing the results of self-attention,channel attention,and convolutional attention integration,CA UNet+++ has the best results on two dalasets with different variation types,LEBEDV and LEVIR-CD,with F1 and OA of 89.2%,97.4%and 87.2%,97.0%,respectively.by analyzing the feature responses of different attention modules to the middle layer of the network,it is demonstrated that the attention mechanism is a potential means to enhance the network potential means of feature extraction capability.(2)To address the problems brought about by the early fusion operation,such as the difficulty in obtaining the high-dimensional features of the original image and the disparity information between image pairs,a change detection network Siamese UNet+++ based on twin structure and full-scale feature reconstruction is proposed to extract the feature information and its disparity information at different scales of the two image phases using the twin structure encoder with shared weights,and in the decoding phase,the disparity information is connected with the high-dimensional information of the later phase In the decoding stage,the full-scale feature reconstruction is performed to reduce the consumption of feature information between layers.The results show that the F1 and OA values of Siamese UNet+++are 96.9%,99.2%and 90.3%,98.3%on LEBEDV and LEVIR-CD datasets,respectively,and the accuracy is improved compared with other comparison methods.Focal dice loss allows the network to focus on the training of difficult samples for classification,thus enabling the network to detect small changes in the image pairs.Focal dice loss is more effective for training Siamese UNet+++with a 3.5%improvement in F1 compared to a mixture of Dice loss and balanced binary cross-entropy combination.(3)The Labelme image annotation tool was used to annotate two scenes of Sentinel2 images of the Red Sand Spring mining area in 2016 and 2020,circle the change area and segmentation using the desertification mining area dataset for validation analysis of the method in this paper.The results show that the CA UNet+++and Siamese UNet+++methods have some applicability on this dataset,and the accuracy reaches the optimal and suboptimal compared with the benchmark methods,with F1 and OA of 80.1%,99.5%and 79.2%,99.5%,respectively,while the above methods are less disturbed by noise than other benchmark methods and can obtain complete and compact change areas of the mine area.