Detection Method Of Glacier Crevasse On Ice Sheet By Sentinel-1 SAR Image Based On Deep Learning

Crevasses,as a typical feature of the surface of glaciers and ice shelves,are of great significance for studying the global greenhouse effect,ice shelf movement trends,and ice shelf stability.It is also the most intuitive factor to directly study the disintegration of ice sheets,andthere a large number of crevasses and hidden crevasses covered by snow,which poses a huge threat to researchers' personal safety.Therefore,the detection of crevasses is very important.At present,the main crevasses detection methods include ground penetrating radar,optical remote sensing,SAR imagery and radar altimetry technology.The SAR image can work all day and all day,and has strong penetrability without being disturbed by clouds and the like.Ithas a strong advantage for crevasses observations.Neural networks can play an important role in the extraction of crevasses because of their strong learning ability and the advantages of being able to apply to complex scenes.U-net network as an encoder-decoder structure can reduce the spatial dimension of the pooling layer,use the existing quick key connection to better repair the target information,and use less data sets to get more precise segmentation results.In this paper,the U-net network is improved,and a new crevasses detection method is proposed,which can be applied to a wide range of scene applications,and the classification results have higher accuracy.The main research results of the paper are as follows:1)Because of the special speckle noise in SAR image,it is necessary to enhance the edge information in the image when the speckle noise is removed in the pre-processing stage,so as to ensure that the fracture edge information can be better preserved and provide clear texture information for later fracture recognition.Based on this,this paper combines MuLOG filtering algorithm and Probabilistic Patch-Based Weights filtering algorithm to process SAR image.This method can effectively retain crevasses edge information and smooth residual noise at the same time2)In order to extract crevasses in the north and south polar ice sheet,it is necessary to process large-scale SAR image,while the traditional extraction method of crevasses only satisfies small-scale extraction,so this paper proposes an improved U-net network.The addition of skip connections in the net network reduces the model training parameters and the calculation time of the network,connects low-level and high-level features,and improves the accuracy of the network to obtain more accurate crevasses segmentation results.In this paper,the improved U-net network is applied to extract the crevasses in the polar region and generate the crevasses product map in the polar region.3)In view of the fact that there are many noise points in the preliminary crevasses detection results,based on the geometric characteristics of the crevasses and the difference in noise,a window traversal method is proposed to perform image post-processing,and the noise points are better removed.4)Taking the Antarctic Larsen C ice shelf crevasses detection as an example,the results of visual interpretation in this area are used to compare and verify the crevasses results extracted by the algorithm in this paper,and the results are judged based on accuracy,precision,recall rate and F1 score.,The results show that a higher accuracy is achieved.