Automated Mapping of Glacier Frontal Retreat in the Indian Himalaya using Satellite Remote Sensing and Deep Learning models

The Himalayan glaciers continue to lose mass due to the undergoing warming climate. They play a critical role in feeding major rivers, such as the Ganga, Indus, and Brahmaputra, supporting the livelihoods of millions. While regional mass changes have been reported by several studies, the current retreat rates are rarely documented. This is primarily due to a lack of satellite data and methods in mapping debris-covered glacier fronts. They are also limited in distinguishing clean ice and perennial snow cover patches. While the global glacier database, such as the RGI (Randolph Glacier Inventory), provides a critical database, it is based on satellite images from 2000-2003.  

In this study, we address challenges in mapping debris-covered glaciers by combining Deep Learning (DL) and a geometric algorithm. We apply several DL models (e.g., including UNet++, GlacierNet-2, GlaViTU, M-LandsNet, and SAU-Net) on multiple remote sensing satellite datasets, which include spectral, radar, topography, geomorphology, and glaciological dynamics. The study sites include four basins of the Himalaya (Chandra Bhaga, Pangong, Chombu Chu, and Alaknanda-Bhagirathi). UNet++ shows the most accurate results with reference outlines, with a mean Intersection over Union (IoU) of ~ 90%. DL-based retreat measurements were closely aligned with those outlined in the reference manual, with a coefficient of determination (R²) of ~ 75%. Our applied Python-based geometric algorithm calculates the average euclidean distance between frontal positions in 2010 and 2019. We find that the retreat rates of debris-covered glaciers in these basins are ~3 m/year during this period.  Lake-terminating glaciers show three times higher retreat rates in the period. This algorithm has the capability to detect glaciers of all sizes, ranging from small to large glaciers, as well as highly debris-covered to clean-ice glaciers, and can identify cirques to hanging glaciers in all the basins.

This DL-based algorithm provides an automated approach with post-processing steps to monitor glacier change with high precision, accounting for the uncertainty of glacier retreat across the Himalaya. This study is important for understanding the relationship between glacier lake expansion and glacier ice mass loss, which can be further used for glacier hazards, such as lake outbursts and dry calving detachments.