High Resolution Remote Sensing Image Change Detection Based On Convolutional Neural Network

Remote sensing image change detection has strong theoretical significance and application value in the fields of land use and cover change monitoring,urban development and change research,environmental monitoring and disaster assessment,etc.,and has always been a hot field of concern for scholars at home and abroad.However,with the improvement of image spatial resolution,high resolution remote sensing images show a lot of new characteristics,such as rich spatial features and object multiscaling,etc.,and the traditional remote sensing image change detection method is limited by artificial design characteristics of the power of expression,on the high-resolution remote sensing image change detection accuracy remains to be improved.In recent years,due to its advantages in deep feature extraction,deep learning technology represented by the convolutional neural networks has shown great potential in the field of image processing.However,the task of remote sensing image change detection has its particularity,which focuses on the change information between images of different times in the same region,so the spatio-temporal correlation between images needs to be modeled to highlight the change information.Therefore,this thesis based on the modeling spatio-temporal correlation between images,from the perspective of based on pixel and based on the object,carried out research based on semantic segmentation models and instance segmentation models of high-resolution remote sensing image change detection method,and further explore the end-to-end change detection network architecture design based on convolution neural network,to achieve higher accuracy and better edge result in high-resolution remote sensing image change detection task.The main work and conclusions of this thesis are summarized as follows:(1)A feature correlation network is proposed to construct the spatio-temporal correlation between images.In this section,a simple component that can model the temporal and spatial correlation is added to the existing typical semantic segmentation model(U-Net)and the instance segmentation model(Mask R-CNN),and combined with the siamese network,so that these models can be applied to solve the problem of remote sensing image change detection.After summarizing the characteristics of U-Net and Mask R-CNN,the Feature Correlation Network(FCo N)is proposed.The FCo N includes two steps of correlation calculation and multi-level feature fusion,which aims to effectively construct the spatio-temporal correlation between images.In the end,we discuss several ways of correlation calculation,and the experiment shows that a simple parameterless correlation calculation method already achieves excellent results.(2)The study on high-resolution remote sensing image change detection under different semantic segmentation models aims to understand the task of high-resolution remote sensing image change detection under semantic segmentation background from the perspective of pixel-based.Specifically,according to the design of FCo N,the three classical semantic segmentation models(Semantic FPN,Deep Labv3,and Deep Labv3+)are modified to change detection models,and relevant experiments were carried out.According to the experimental results,a detailed quantitative and qualitative analysis was conducted,and the advantages and disadvantages of these models are also analyzed.According to the experimental results and comparative analysis,for change detection model’s network architecture design and performance based on semantic segmentation,we can draw the following conclusions: 1)Multi-level feature fusion can improve the performance of the model,and the more hierarchical features in the feature extractor,the better the performance and generalization ability of the model,which also proves the necessity of multi-level feature fusion in the FCo N;2)The prediction of the selected typical semantic segmentation model on the edge is still far from the ideal result.Especially for the dense change area,these models are easy to predict the change area of multiple areas into a area,which indicates that the edge prediction ability of the model needs to be improved.(3)The study on high-resolution remote sensing image change detection under different instance segmentation models aims to understand the task of high-resolution remote sensing image change detection under the background of instance segmentation from the object-based perspective.Specifically,according to the idea of FCo N,the two typical instance segmentation models(Blend Mask and Center Mask)are modified into change detection models,and relevant experiments were carried out.According to the experimental results,a detailed quantitative and qualitative analysis was conducted.Finally,the advantages and disadvantages of the change detection model based on instance segmentation model(Mask R-CNN,Blend Mask,and Center Mask)and the change detection model based on semantic segmentation models(UNet,Semantic FPN,Deep Labv3,and Deep Labv3+)were compared and analyzed.According to the experimental results and comparative analysis,the main differences between the change detection method based on the instance segmentation and the change detection model based on semantic segmentation are as follows: The change detection models based on instance segmentation are to segment the sub-region of the image,while the change detection model based on semantic segmentation are to segment the entire region of the image,therefore,the change detection models based on instance segmentation,are better at separating the dense change areas,but is easier to divide a large area into multiple areas,and change detection models based on semantic segmentation are just the opposite.(4)The Edge Sensing Network(ESNet)is proposed to obtain better change detection accuracy and effectively retain the detailed information of the edge.ESNet is composed of a backbone,coarse prediction branch,and fine prediction branch.The backbone is used to extract multi-level correlation features.The coarse prediction branch is used for the fusion of multi-level feature maps to model the consistency of change object’s inner.The fine prediction branch is used for the fusion of multi-level feature points to retain edge details as much as possible.Compared with other methods,the ESNet can not only obtain the optimal change detection accuracy but also effectively retain the detailed information of the edge.