Automatic Detection Of Himalayan Glacial Lakes Using Deep Convolutional Neural Network

Glacial lakes are important triggering factors for glacier melting,floods and debris flows,and are also important reference samples for studying global warming.Usually,the collapse of a glacial lake will cause a regional collapse,often to the middle and lower reaches,and its water volume will exceed the warning water level of the year,and the safety of people's lives and properties in the downstream will be affected by the collapse of the glacial lake.Traditional glacial lake research relies on artificial method to complete the interpretation of the target,which is time-consuming and labor-intensive.Therefore,it is of great significance to use deep learning to find exactly where the ice lake is.This paper takes the ice lake image in Sentinel-2 remote sensing image as the research object,and realizes the instance segmentation of ice lake image based on the improved Cascade R-CNN algorithm.The main research contents and conclusions of this paper are:(1)Instance segmentation of glacial lake images based on lightweight network.An instance segmentation model of glacial lake images based on HRNet,HL-HRNet,is proposed.In Cascade R-CNN,Res Net is used as the Backbone,and its feature map may reduce the resolution,and the feature expression ability is limited.Learn the importance of feature channels by changing the convolution block of the Stem layer in HRNet to lightweight Shuffle Net,and notice that the 3*3 convolution in Shuffle Net increases the number of parameters and splits it into 3*1 and 1* The combination of 3 convolutions increases the learning weight of effective features for ice lake segmentation for HRNet,suppresses other unimportant features,and improves the ability to extract features.(2)Glacial lake image segmentation based on deconvolution algorithm.Aiming at the problems of low accuracy and high missed detection rate of HRNet instance segmentation model in complex scenes,an improved HRNet algorithm based on deconvolution is proposed.Improve the detection and segmentation ability of HRNet algorithm for small targets and the missed detection and false detection caused by ice lake freezing,water pollution,shadow occlusion,etc.in complex scenes.The improved HRNet has obvious advantages in solving the occlusion problem in complex scenes,and has better segmentation performance.Experimental results show that Cascade R-CNN based on improved HRNet has higher segmentation performance.Compared with the original Cascade R-CNN using Res Net as the backbone network,the recall rate is improved by 6.4%,the accuracy increased by 2%,and the overall map increased by3.7%.(3)In this study,this paper obtained the world's first 10m-resolution Himalayan glacial lake data inventory,filling the gap in the study of glacial lakes under 10 m resolution images in the Himalayas,which revealed the current spatial distribution of Himalayan glacial lakes in 2020,with a total of 4,950 previous Himalayan glacial lakes with the highest precision datasets.The total area of glacial lakes is about 455.2855 square kilometers.In the Himalayas in the same study area,the results of this paper include 8831 glacial lakes totaling 903.6865 square kilometers,an increase of 3881 lakes and an increase of 448.4010 square kilometers compared to the 2015 estimate.km,the number of ice lakes increased by 98.5%.This study can serve as a baseline data source for future scientific research,such as assessing changes in Himalayan glacial lakes.