The Research Of Detection And Risk Evaluation Of Glacial Lakes Based On Feature-Learning Methods Across High Mountain Asia

Under the influence of continuous global warming,phenomena such as melting glaciers,increasing precipitation,thawing permafrost and shifting of permanent snow line have occurred in the cryosphere.Glacial lakes formed by pooling glacial meltwater in the depression.In High Mountain Asia(HMA),glacial lakes are developed from east to west in Nyainqentanglha,Himalayas,Tibet Plateau,Pamir Plateau,Kunlun Mountains,Tienshan Mountains,etc.With the expansion of the glacial lake area,dam failure occurs when the moraine dam cannot suffer the stress from the lake water,concomitantly causing Glacial Lake Outburst Flood(GLOF)disaster and posing a great threaten to the life,infrastructure of the downstream area.Therefore,monitoring the dynamics of glacial lake,summarizing the principle of glacial lake outburst,and assessing the potential outburst risk of glacial lake are of great importance for exploring the relationship between glacial lake changes and global climate change,and carrying out disaster prevention and reduction of glacial lake in plateau areas.Glacial lakes are mostly developed in the plateau areas where has harsh natural conditions and few people,it is difficult to conduct a large-scale field investigation of glacial lakes.Remote sensing technology has the advantage of large-scale remote observation of the surface,and has made many achievements in the research of glacial lake identification,extraction and monitoring.However,due to the limitation of computer computing power,most of these achievements are extracted and analyzed by manually designed features of glacial lakes in a regional scale.In general,auxiliary data,such as Digital Elevation Model(DEM),are still needed for pre-and post-processing.As a result,the efficiency of large-scale glacial lake monitoring is still low and is difficult to summarize the general rules of glacial lake outburst.In recent years,with the development of computer vision,image processing methods based on deep learning have made great progress in many image-related tasks by using convolution and other operations to extract high-level semantic information in images and mine potential relationships between objects.Therefore,this study combined the characteristics of glacial lakes and relevant technologies of deep learning to mine glacial lake information in remote sensing images and carry out a series of designation on glacial lake outburst,including the selection of spectral features of glacial lake,the designation of glacial lake extraction algorithm,the analysis of glacial lake change,and the modeling of glacial lake outburst,so as to realize the rapid monitoring and outburst risk assessment of glacial lake in the High Mountain Asia.This research has obtained the following results and conclusions:1.Systematically analyzing and evaluating the optimal spectral features of large-scale glacial lake monitoring.In this experiment,23 common spectral features of water bodies were selected.Then random forest and decision tree were used to score 23spectral features,and some useful features were selected based on the quantitative analysis and visualization results,Normalized Difference Water Index(NDWI)obtained the highest kappa coefficient(0.8812)and prediction accuracy(0.9025),which was much higher than other spectral features.The results show that the spectral feature is more suitable for large-scale and multi-types of glacial lake extraction,and provides a basis for the subsequent glacial lake extraction work.2.Designing a generative adversarial network for glacial lake(GAN-GL)extraction via a supervised training fashion.This work was based on Generative adversarial Network(GAN),and designed an automatic extraction model of glacial lakes without pre-processing,post-processing or auxiliary data de-noising.The model contains a water attention module(introduced NDWI),image segmentation module,and the discriminator with Res Net-152 structure.The loss term is composed of adversarial loss and content loss.Through evaluating on glacial lake extraction on Landsat data,and comparing with glacial lake iterative segmentation model and random forest model,the results showed that the proposed model GAN-GL could effectively avoid the interference of melting glaciers and mountain shadows and accurately extract glacial lakes in the eastern Himalayan region.The F1 score(0.7331)and Io U(0.5846)were much higher than those of iterative segmentation model(F1=0.3681,Io U=0.2266)and random forest model(F1=0.5363,Io U=0.3584),and without use of auxiliary data,pre-and post-processing,which showed great potential of large-scale automatic extraction of glacial lake.3.Proposing a simple glacial lake(Sim GL)extraction model via contrastive learning and weakly-supervised training fashion.This experiment combined with the idea of contrast learning,proposed a simple glacial lake(Sim GL)extraction model without making true labels of glacial lake.The model consists of two modules,namely,the comparison module between the input image and the transform image and the position module between the input image and the NDWI image results.The loss function is composed of the contrast loss and the position loss.Evaluation results on Landsat-8 OLI dataset showed that the proposed method is better than the traditional glacial lake extraction model(F1=0.6661,Io U=0.5855),and can achieve the considerable effect with the supervised model(such as random forest F1=0.6649,Io U=0.5634).By testing the model Sim GL on different series of Landsat images and Sentinel-2A images,it is also proved that the model has good applicability and is more suitable for different remote sensing data of glacial lake extraction.4.Assessing the outburst risks and changes of glacial lakes across High Mountain Asia in the past 30 years.In this experiment,350 glacial lakes with an area greater than0.1km~2 and a change area greater than 5000m~2 in High Mountain Asia in recent 30 years were filttered.Combined with remote sensing images,elevation data,meteorological data and administrative data,the multi-factor outburst risk assessment and visualization analysis was carried out for 350 glacial lakes with high outburst risk.The experimental results showed that in the past 30 years,Ganglongco(85.82°E,28.32°N)and Gangxicco(85.87°E,28.36°N)are the two glacial lakes with the highest risk in high Asia,and their outburst risk coefficients are 0.904 and 0.861,respectively.Nyalam County lies 20km downstream to the southwest.It is necessary to make continuous observations of the two glacial lakes.This work assesses the risk coefficient of glacial lake in the whole region of High Mountain Asia,and provides the basis and guides significance for disaster prevention and reduction in the downstream area.