Predicting Mass Balance of Glaciers in the Central Himalaya Region using Machine Learning

Abstract

Glaciers serve as critical indicators of climate change worldwide, and with rapid glacier melt and increased vulnerability to Glacial Lake Outburst Floods (GLOFs), it is necessary to assess regions like High Mountain Asia (HMA), which hosts a dense concentration of glaciers and glacial lakes, for glacier health. One key metric is the Glacier Mass Balance (GMB), which represents the net mass gain or loss of a glacier over a given year. While traditional temperature-index-based machine learning models have been developed for GMB prediction, the influence of energy balance and solar radiation—especially at low latitudes—remains underexplored. This study predicts the GMB for low-latitude HMA glaciers using machine learning models, including Linear models, XGBoost, and Multi-Layer Perceptron (MLP). The dataset for the model includes key feature groups, including topographic, meteorological, and radiation. The target variable for training and evaluation was derived from the Hugonnet glacier mass balance dataset (2000–2019). To prevent data leakage and ensure spatial robustness, glacier-aware dataset splitting methods such as StratifiedGroupKFold were used. Among the models tested, XGBoost performs best with an R² of 0.559 and RMSE of 0.259, followed by MLP (R² = 0.0.512 and RMSE 0.273). Linear models (Ridge and Lasso) perform the worst, with similar R² values of 0.335 and RMSE of 0.319. Feature importance and SHAP value analysis reveal that latitude, longitude, and topological features are dominant predictors, highlighting the role of geographic location in glacier dynamics, mainly in low-latitude Himalayan regions. Overall, this study demonstrates the use of solar radiation and climate features for predicting GMB and investigates their role in the overall influence on GMB prediction at the low-latitude HMA glacier.