Change Detection Based On Encoder Decoder For High-Resolution Remote Sensing Image

With the rapid development of the economy and society and the advancement of urbanization,the area of cities and towns has continued to increase,and land use & land cover(LULC)information is undergoing rapid changes.Information on the changes in coverage has important roles and significance in practical areas such as urban development planning,land management,and vegetation cover.As an important data source for ground observation,high-resolution remote sensing images have been widely used in surface cover information monitoring.After the remote sensing image data is interpreted,it not only provides solid and reliable data support for decisionmaking departments,but also brings up the problem of how to quickly and effectively process a large number of remote sensing images.In order to extract the change information of remote sensing images using traditional algorithms,artificially designed image features need to be extracted,and the model generalization ability is not strong,and the entire change detection process requires human intervention,tedious processes,and low degree of automation.This paper uses semantic segmentation in deep learning.The problem of remote sensing image change detection is transformed into a variable and unchanged binary segmentation problem.The end-to-end full convolutional neural network is used to detect changes in the remote sensing image before and after two periods.The main research contents of this paper are as follows:(1)Change detection of remote sensing images based on the encoding-decoding structure.In order to introduce multi-resolution information of the two periods before and after the image,a center surround module is introduced in the change detection network model structure,and the change detection is replaced by a convolution layer with a step size of 2.The pooling layer in the network to prevent the loss of image information due to the pooling process.ASPP module is added to the change detection network to introduce more image context information.In addition,the shallow and deep features are fused attention,the introduction of attention gating mechanism to improve the accuracy of change detection.(2)In order to compensate for the easily neglected pixel-to-pixel relationship based on the result of the encoding-decoding structure change detection network,which affects the change detection result,a generative adversarial network is introduced.The change detection network based on the encoding-decoding structure is used as the generation network of the GAN network,and the Wasserstein distance is used to replace the JS divergence in the original GAN network model.The proposed method can improve the accuracy of change detection in remote sensing images.This thesis has 36 figures,6 tables and 116 references.