Modelling And Methodology Based On Deep Learnining For Extracting Spatiotemporal Change Information Of Alpine Glacial Lakes By Using Optical And SAR Images

As an important part of mountain lakes,alpine glacial lakes are an important link connecting the hydrosphere and cryosphere.In my country's Qinghai-Tibet Plateau,with the gradual warming of the climate,the melting of glaciers has intensified,leading to frequent occurrences of glacial lake outbursts.Efficiently and accurately extracting the contours of alpine glacial lakes and analyzing their temporal and spatial changes are of great significance to the evaluation and prevention of glacial lake outburst disasters.Remote sensing is a non-contact remote imaging technology.Thanks to its advantages of wide coverage and high imaging resolution,it has been widely used in the extraction and analysis of ice lakes since the last century.However,the use of remote sensing images to extract ice lakes currently has three main problems:(1)Optical remote sensing images are susceptible to cloud and fog interference.Although SAR images can penetrate clouds and fog,they are prone to coherent noise,and due to the SAR side-view imaging mechanism,there will be Severe geometric distortion results in low accuracy of ice lake extraction;(2)The imaging mechanism of optical image and SAR image is different,and there is a certain geometric deviation of the pixels between the images,so it is difficult to combine the two remote sensing images for ice extraction.Lake extraction;(3)The distribution of glacial lakes is generally discrete,and the number of glacial lakes contained in each remote sensing image is often large,and the efficiency of manually labeling glacial lakes is low.At present,the automatic extraction technology of ice lake based on remote sensing images mainly focuses on the grayscale features of pixels,or the spatial features such as artificially designed aspect ratio and roundness.The classification accuracy is low,and the robustness of the extraction scheme is poor,and it cannot be applied to different types.Multi-scene ice lake extraction in regions and different weathers.In order to overcome these problems,this paper systematically studied the spectral and texture characteristics of ice lakes,as well as the advantages and disadvantages of optical and SAR images,tested and compared the effectiveness and accuracy of various image classification methods applied to ice lake extraction,and proposed the use of convolution The neural network establishes a PPC-Unet(Pyraimd Pooling Cascade-Unet)model,which aims to fuse the characteristics of optical-SAR heterogeneous remote sensing images to accurately extract ice lakes.In view of the large terrain undulations in the study area,and the geometric distortion of SAR images have a greater impact on the extraction of ice lakes,this paper proposes to use the method of combining ascending and descending orbit images to remove mountain shadows.According to the characteristics of the reflection signal of the ice lake,the maximum value of the pixel gray value of the ascending and descending orbit image is used as the target value,and the gray value of all pixels greater than 1 is set to 1,which is more in line with the PPC-Unet model.In order to reduce the image migration error,this paper also uses mutual information method to screen the registration points and geometrically correct the7 bands of Landsat 8 image and Sentinel 1 image one by one.In view of the discrete distribution of glacial lakes and the large amount of remote sensing data,the existing automated extraction methods are difficult to accurately extract glacial lakes.This paper uses a high-accuracy deep learning method,and performs feature extraction,assigning weights,and weights based on the features of the two images.Feature fusion,thereby designing the PPC-Unet convolutional neural network model,and obtaining robust extraction results.In order to verify the accuracy of the glacial lake extraction method proposed in this paper,this paper takes the southeastern Tibet as the research area to carry out glacial lake extraction experiments.The research results show that: using the same deep learning model,the combined optical and SAR remote sensing image ice lake extraction effect is significantly better than the extraction results of the optical remote sensing image or the SAR remote sensing image alone,and the deep learning method can better integrate the two heterogeneous sources.Advantages of remote sensing images: There are a large number of glacial lakes in southeast Tibet.According to the method of glacial lake extraction proposed in this paper,a total of 7884 glacial lakes have been obtained,of which about 7,200 are stably existing glacial lakes,and 1,224 are glacial lakes with a larger area;The area of the lake gradually increases,and the accuracy of PPC-Unet's extraction of ice lakes gradually increases.The area of glacial lakes has been increasing during 2015-2019,and the overall change is relatively slow.Among them,the total area of glacial lakes has increased significantly from 2015 to 2016,and the area of glacial lakes has expanded in 2018-2019.Analysis shows that changes in the area of glacial lakes are not only affected by glacier melt water,but also by short-term heavy precipitation.The area of glacial lakes close to glaciers increases more rapidly.There are many differences between the imaging mechanism and image features of optical and SAR remote sensing images.By using the deep learning method to automatically assign the weights of the two images in different environments,the features of the two heterogeneous images can be effectively merged to achieve high Accurately extract the boundary information of the ice lake.The research results in this paper can provide key technical support for the extraction of alpine glacial lakes and the prevention and control of glacial lake disasters.