Glacier Extraction Of Single-polarization SAR Image Based On Multi-scale Joint Convolution Neural Network

Large-scale change of glacier area is one of the important conspicuous indicators of regional and global climate and environment change.Synthetic Aperture Radar(SAR)as a means of obtaining ground information that is not affected by time,climate and environment,have attracted much attention in the field of glacier monitoring.Glacier extraction from images is an important step in glacier change monitoring.As the mainstream means of glacier monitoring at present,U-Net has been able to achieve efficient and accurate glacier extraction from full-polarization and dual-polarization.However,there are still some shortcomings in long term glacier monitoring from single-polarization SAR images.First of all,the single-polarization SAR image has less feature information and lower dimension.In the area of snowline,it is difficult for U-Net to accurately identify snowline by using features extracted from shallow convolution structure.Secondly,the glacier boundary has a lot of detailed information with irregular shape,and the resolution ability of U-Net for detailed information is limited by the receptor field of the fixed convolutional kernel,so the retained glacier boundary details are less.Finally,the glacier extraction results contain noise points and external points,and the current glacier extraction algorithms lack effective optimization methods.In order to solve the above problems,this paper proposes a glacier extraction strategy based on multi-scale joint convolutional neural network combined with conditional random field post-processing.The main work of this paper is as follows:(1)Based on the U-NET network structure,the segmentation network is designed to be more suitable for glacier extraction.This paper mainly improves from the following two aspects.First of all,the subsampling network is reconstructed with the residual structure of the deep separable atrous convolution as the convolutional layer.The improved subsampling network can effectively improve the ability of deep feature mining,which can not only accurately judge the position of the snow line,but also prevent the occurrence of degradation problems without introducing too many parameters.Secondly,multi-scale joint convolution is established to extracted global and local information.Different scales atrous convolution kernels and global average pooling are used to extract the information of images with different resolutions.In this way,the receptive field of the network is no longer limited to the fixed size of the convolution kernel.The accuracy of the network for the glacier boundary is further improved through feature fusion combines global and local information.(2)Design the optimization algorithm of glacier classification post-processing.A classification post-processing algorithm based on Conditional Random Field(CRF)model is designed and implemented.A conditional random field is constructed by using the classification probability map extracted from the glacier network,and then the noise points and local points in the classification results are searched and corrected by the potential function.(3)Verify the glacier extraction algorithm.In this paper,Taku Glacier is taken as the extraction target,and Terra SAR-X(TSX)satellite image is taken as the experimental data.Training and validation data sets were constructed through visual interpretation.The glacier extraction algorithm in this paper is compared with the commonly used Support Vector Machine(SVM),U-Net glacier extraction network and the image post-processing optimization algorithm respectively to verify the effectiveness of the proposed method.What's more the Sentinel-1A(S-1A)satellite images representing different sensors,polarization mode acquisition times and resolutions are used for transfer training to verify the transfer learning ability of the proposed method.To comprehensive quantitative validates the effectiveness of the proposed algorithm,the accuracy,intersection over union,recall,specificity,precision and F-1 value were used to quantitatively evaluate the experimental results.And,the experimental results were qualitatively evaluated by visual interpretation.The experimental results show that the proposed glacier extraction algorithm can effectively extract the glaciers in the large range of remote sensing image area.