Study On Segmentation Algorithm Of Glacier Regions Based On Remote Sensing Images

China has a large number of mountain glaciers,which have dramatic effects on the economic life of the surrounding areas.Therefore,the research on glaciers has great practical value.Monitoring Glacier is one of the important means of glacier research.By research the area and volume change of glacier,researcher can understand the change information of ecological environment.With the development of semantic segmentation technology,geographic researcher uses the networks of semantic segmentation to complete the task of remote sensing images glacier segmentation.Based on the knowledge of semantic segmentation,we apply the semantic segmentation network to complete the task of remote sensing images glacier segmentation and make improvements to the network to improve the effect of the whole network.The main contents of this dissertation are as follows:(1)Because of losing glacier segmentation datasets,this dissertation uses multi band remote sensing images and geographic vector files to make glacier datasets and expands the existing datasets with data enhancement method.Using multi band remote sensing image data improves the effect of the whole network.(2)Aiming at the problem of low segmentation accuracy of traditional glacier segmentation methods in this kind of glacier segmentation task,a glacier segmentation method based on multi band remote sensing image is proposed.In this dissertation,the glacier segmentation task is carried out through the classical deep learning semantic segmentation network u-net.By adding the band combination and enriching the spectral information,the problem of segmentation accuracy of traditional glacier segmentation methods is solved.By adding the module of dilated convolution and spatial attention,the experimental accuracy of the whole network is improved.Through the test of the network,it is found that the u-net proposed in this dissertation has achieved 93.1% and 85.1% in Mean Pixel Accuracy and Mean Intersection over Union,which is 1.7% and 0.7% higher than the original u-net network.Among all networks,the u-net proposed in this dissertation is the best.(3)Aiming at the problem which is difficult to distinguish glacier and water by semantic segmentation network,a semantic segmentation network to distinguish glacier and water is proposed.This dissertation proposes the network structure of u-net,and use depth-wise separable convolution,the shuffle attention to improve the effect of the network.Through the depth-wise separable convolution,the calculation of the network is reduced.Through the shuffle attention,while adding a small amount of calculation,the effect of the network has been improved.The experimental results show that the proposed network has a certain improvement in the evaluation compared with the original u-net.The Mean Pixel Accuracy is91.4% and the Mean Intersection over Union is 82.6%.The network improves the segmentation accuracy to a certain extent.