Research On Remote Sensing Image Change Detection Method Based On Deep Learning

Remote sensing image change detection has attracted widespread attention and importance because of its ability to monitor surface changes in a large area and to understand the development trend of surface changes comprehensively and quickly.It can precisely locate the area of change and identify the type of change.According to the different degrees of decoding of the identified change information,the remote sensing image change detection problems can be divided into two categories: binary change detection problems and semantic change detection problems.In recent years,with the increasing maturity of computer technology,deep learning technology has been organically combined with remote sensing image change detection.However,there are still some limitations of deep learning techniques in remote sensing image change detection tasks.For binary change detection,due to the complexity of feature information in remote sensing images and the differences in sensor imaging conditions,features of the same scene present different expressions at different times and different spatial locations,which makes it difficult to effectively identify the pseudo-change regions in remote sensing images.In addition,the serious imbalance between changed and unchanged data in remote sensing image data samples directly leads to the low accuracy of remote sensing image change detection results.For semantic change detection,the proliferation of high spatial resolution and high temporal resolution sensors cannot be fully matched with equally rich labels,which can lead to the inability of semantic change detection to train a change detection model with excellent performance in a supervised manner.In addition,deep convolutional networks still have some problems in semantic change detection tasks,such as difficult interpretation and small perceptual fields.Therefore,this paper uses deep learning techniques to optimize two remote sensing image change detection methods with different degrees of interpretation,respectively,and the main research is as follows:(1)To address the problems in remote sensing image binary change detection,we first extract high-level image semantic features from remote sensing images using the twin Efficient Net V2 network,and transform the dual-temporal remote sensing image feature maps into compact semantic token sets using the spatial attention mechanism,respectively.Based on this,the context within the two semantic tag sets is modeled using the Transformer encoder,and the semantic tags are reprojected to pixel space using the Transformer decoder to enhance the original pixel-level feature representation by exploiting the relationship between pixels and semantic tags.Finally,Dice Loss and Focal Loss are summed to generate a new loss function to alleviate the problem of sample class imbalance.The model proposed in this paper was compared with five other methods for experiments,and the results show that the method in this paper can distinguish real changes and complex uncorrelated changes from remote sensing images more accurately,thus improving the accuracy of identifying feature change information.(2)To address the problems in the semantic change detection of remote sensing images,we propose to use the label super-resolution method based on epitomes to replace high-resolution labels with noisy,low-resolution weak labels for semantic change detection,from which changes in specific feature type information are detected.Firstly,low-resolution satellite data are used to smooth the quality difference problem of high-resolution remote sensing image input,and color enhancement is applied to the input remote sensing data;secondly,the high-resolution remote sensing image classification map is predicted by combining the label super-resolution algorithm as a statistical inference algorithm and the epitomes model;Finally,a small full convolutional network is fitted to post-process the generated remote sensing image classification map to improve its classification.The model proposed in this paper was compared with five other methods for experiments,and the results show that the model proposed in this paper exhibits superior semantic change detection capability and can effectively detect the specific types of feature change information.