Research On Glacier Extraction Algorithms Based On Data-driven Deep Learning

As an important freshwater resource in the world,glaciers play an important role in the monitoring of the earth's environment,especially in the study of global climate change.First of all,glaciers are very sensitive to regional and global climate change.Climate change leads to changes in temperature and snowfall,changes in the balance of surface quality,and changes in glacier area.Secondly,the global sea level changes affected by glaciers.Over the past century,glacial melting has been the main cause of global sea-level rise.In addition,glaciers are closely related to human production and life,and play an important role in water storage and runoff regulation.Satellite monitoring technology,especially multispectral satellite,provides rich spectral information for the monitoring of ground targets,and has great advantages for glacier monitoring.The Sentinel-2 satellite,which is launched by ESA,has the characteristics of high spatial resolution,short revisit period and free user access,and has attracted the attention of scholars.In order to study the potential of sentinel-2 imagery for glacier monitoring,the main work that has been done is described as following:(1)According to the Sentinel-2 imagery,which is rich in spectral information and has spatial correlation between ground objects,this paper proposes a spatial spectral module,which can establish band or spatial correlation by suppressing or highlighting the spatial and spectral information in the image,so as to make more effective use of sentinel-2 image information to extract glaciers and reduce environmental impact ?(2)In order to achieve faster and more accurate glacier extraction,this paper makes adjustment and Optimization for the structure of UNet neural network,discusses the best position of spatial spectrum module in UNet neural network to improve the classification accuracy.In order to prove the superiority of the proposed model,it is compared with the traditional glacier classification method and classical neural network.(3)According to the generalization ability of the model,this paper evaluates the stability of the improved UNet neural network with time change,uses different time interval training sets and test sets to quantify the time correlation of the model and evaluate its test accuracy.In this paper,the Sentinel-2 imagery is used to analyze the glacier extraction algorithm,and a glacier extraction algorithm based on deep neural network is proposed.The results show that the proposed neural network has higher accuracy and stability for glacier recognition.