Detection Of Antarctic Surface Meltwater Via Deep Learning

The Antarctic plays a crucial role in global change research due to its immense size and its influence on ocean and atmospheric systems.The changes occurring in the Antarctic ice sheet not only have significant implications for polar ecosystems but also for the global climate system and sea level rise.The Antarctic ice sheet holds a vast amount of the earth’s freshwater resources,and if it were to melt entirely,the coastal regions where human activities currently exist would be submerged,leading to catastrophic consequences.Even if the ice sheet were to melt partially,it would have a severe impact on sea level rise,and accurate predictions are essential to plan and develop policies to mitigate its impacts.Currently,the Antarctic ice sheet poses the greatest uncertainty to global sea level rise,and researchers have limited knowledge of the drivers of glacier dynamics and mass balance affecting it.Hence,studying the Antarctic ice sheet and its changes is vital to understanding the impacts of climate change and to implement measures to mitigate its consequences.Surface meltwater is a crucial component in understanding the dynamics of the Antarctic ice sheet.It is formed through melting at higher elevations and gradually converges towards the lower elevations of the ice shelf.As the formation of surface meltwater is one of the main forms of response to climate change and one of the main forms of Antarctic mass loss,it is important to understand its impact.Surface meltwater affects the Antarctic mass balance in three main ways.First,when surface meltwater enters the base of the ice sheet through fractures in the surface of the ice cap or ice shelf,it connects to the hydrological system at the base of the ice shelf,accelerating the basal sliding of the ice sheet.This in turn accelerates ice flow movement and contributes to Antarctic mass loss.Second,surface meltwater forms runoff on the Antarctic surface,discharging Antarctic mass directly into the ocean,which causes mass loss.The freshwater discharged into the ocean also has an important impact on ocean circulation.Finally,surface meltwater forms numerous supraglacial lakes on the ice shelf surface,and the formation and drainage of these lakes can cause ice shelf deformation by accelerating basal sliding,directly causing mass loss through runoff discharge into the ocean,and triggering ice shelf collapse.The formation of supraglacial lakes and their drainage can also cause ice shelf deformation and subsequent collapse.Collapse leads to lower back stresses of the ice shelf and thus faster ice flow,which increases Antarctic mass loss.As such,understanding the dynamics of Antarctic surface meltwater is crucial in accurately predicting the future of the Antarctic ice sheet and its impact on global sea level rise.Remote sensing satellite imagery is a useful tool for studying surface meltwater in Antarctic.However,the wide distribution of surface meltwater and the large variation in lake size make manual visual classification time-consuming and labor-intensive.To automatically extract surface water bodies from Sentinel-2 multispectral imagery of the Antarctic,this paper employed a deep learning approach.There are two main challenges in applying deep learning to the identification of surface water bodies in Antarctica:first,how to construct high-quality training data,and second,how to select an appropriate deep learning model that can effectively utilize the characteristics of the corresponding satellite imagery.This paper proposes a feasible workflow for addressing these two challenges.In addition,to validate the application of the model to the study of surface water body evolution in the research area,this paper analyzes the spatio-temporal changes of surface water bodies on the Amery Ice Shelf from 2017 to 2022.The main contents of this article are as follows.First,the Normalized Difference Water Index(NDWI)was used to construct the training and testing sets for deep learning.NDWI uses the band information of different channels in multi-spectral satellite images to extract surface water bodies through a threshold method.It is a relatively mature method that has been successfully applied in the Southwest Antarctic and Greenland ice sheet.Based on NDWI,masks for rocks and clouds are added to remove the influence of interfering features.However,false positives still exist in the extraction results of NDWI,requiring a small amount of manual intervention to establish a high reliability dataset.Second,the multi-channel and high spatio-temporal resolution features of Sentinel-2 was utilized to add an attention module to the U-Net model for automatic extraction of surface water bodies in the Antarctic.The attention mechanism assigns different weights to different features to increase the U-Net model’s focus on the region of interest and improve its recognition accuracy.Comparing the recognition results of the original U-Net model and the U-Net model with the attention module,the modified U-Net model has higher accuracy and Kappa coefficient,with an average accuracy of 0.9969 and a Kappa coefficient of 0.9018.Finally,the spatio-temporal changes of surface meltwater on the Amery Ice Shelf from 2017 to 2022 was analyzed.Spatially,surface water is mainly concentrated in the inland area of the Amery Ice Shelf between 70-73°S,accounting for 96%of the total area.Within 20 km of the grounding line of the Amery Ice Shelf,water bodies account for 93%of the total area.Temporally,the water area changed significantly in different years,with the maximum area occurring in 2017,reaching 932.54 km~2,and the minimum area of only 58.35 km~2 occurring in 2021.The spatial distribution of surface water on the Amery Ice Shelf is controlled by factors such as firn air content,katabatic winds,exposed rock,and blue ice,while the interannual variation in surface water area on the Amery Ice Shelf is associated with complex climate factors,including temperature,surface net solar radiation,snowfall,and snowmelt,among which temperature and snowfall show strong correlation.