Research On Remote Sensing Image Change Detection Algorithm Based On Convolutional Neural Network

The change detection of remote sensing images refers to shooting the ground surface through a series of high and low altitude remote sensing platforms such as satellites and unmanned aerial vehicles,and obtaining images for detection to locate the changed areas and identify different types of changes.Change detection using remote sensing images is of great significance for guiding geological exploration,resource management,and environmental monitoring.Compared with traditional change detection methods,deep learning detection methods are more accurate and robust.Therefore,this paper studies remote sensing images from the perspective of two-category change detection and multicategory change detection.The main research work is as follows:(1)Aiming at the problems that the UNet++ network directly upsamples the extracted features without fully extracting the semantic information of the image and utilizes information sharing between layers,a binary change detection method based on ZUNet++ is proposed for remote sensing images.The ZUNet++ network can extract more refined semantic features of remote sensing images by adjusting the dense skip connections between UNet++ nodes,again reducing the semantic gap between the feature maps of the encoder and decoder.Secondly,the change detection map of remote sensing image is obtained by introducing fuzzy C-means clustering.Furthermore,to address the lack of multispectral datasets,a recombination strategy that converts a single multispectral image into multiple pseudo-RGB images is proposed to enhance the dataset.Experiments show that by comparing with FCN,UNet,and UNet++,the proposed method can detect more complete change regions,and the model is less affected by noise points.(2)Aiming at the problems of blurred edge contour and long training time in the detection results of STANet for small buildings,a binary change detection method MSNest Net based on multi-scale nested Siamese network was proposed.The features of single-phase images are first extracted using a residual network that introduces efficient channel attention.Secondly,the two groups of different features extracted are mapped to the same feature space,and the multi-scale nested model is used to extract the fused deep semantic feature information for similarity judgment,so that the correlation of unchanged pixels increases and the correlation of changed pixels decreases.Finally,the problem of class imbalance is solved by the Focalloss loss function.The experimental results show that,compared with STANet,Siam-Nested UNet,and DSMSCNNet,the proposed method is more complete for edge detail detection of small buildings.(3)Currently,compared with two-category change detection,less research has been done on multi-category change detection.We propose a staged approach for multi-class change detection.The first stage fully extracts the global information of remote sensing images through the cross-fused high-resolution network HRNet;the second stage uses the principal component analysis mean clustering(PCA＿Kmeans)to obtain the result map of multi-category change detection.The experimental results show that the proposed method is relatively complete in detecting the contour lines of buildings with various types of changes,and has a lower error rate.