Automated Alpine Glacier Mapping Using Deep Learning Approach

Abstract:

In recent decades, the glacier behavior, dynamics, and processes are affected by climate change. Since mountain glaciers such as those in the Himalaya and Karakoram are the source of many river systems, it is essential to quantify these behavioral changes. One such seemingly simple but notoriously difficult aspect is accurate and precise glacier mapping, especially for the complex Himalayan debris-covered glaciers. Although many methods have been developed to overcome this issue, automated mapping remains a challenge. Therefore, in this study, an advanced deep learning-based approach is developed for automated glacier mapping. The approach takes advantage of the convolutional neural network (CNN) to delineate the ablation zone. To classify the accumulation zone, we utilize the drainage basin algorithm that estimates the drainage basin based on the topography data. The input data includes Landsat satellite imageries, digital elevation model (DEM), and geomorphometric data. Training data labels were generated using glacier shapefiles downloaded and modified from an online database, such as the Global Land Ice Measurements from Space (GLIMS) and GAMDAM glacier inventory. Because of the large swath of the input data, we subsample them into several sub-images to fit the CNN input requirement. The CNN architecture used in this approach is the segmentation model that is designed for semantic segmentation. This model extracts and classifies the features from the input data, and produces a binary image to label all the target ablation zone pixels. The drainage basin algorithm identifies all the snow-covered accumulation zone by using a normalized difference of snow index, and DEM to find the pixels that flow into the ablation zone. We tested our methods on the Himalayan glaciers and compared it with the online database. Our preliminary results show high-quality performance in delineating retreating, advancing or stable terminus; debris cover glacier ablation zone; snow-covered accumulation zone; and nearby mountain ridges. The overall accuracy in mapping the ablation zone is higher than 0.8 intersections over union after learning 10-20% of the total study area.