Study of Himalayan Glacier Mass Balance Processes using Machine Learning

Abstract

Glacier mass balance is an important indicator of climate change and a valuable metric in understanding the vulnerabilities of ecosystems and societies to future climate scenarios. Approaches to modelling glacier mass balance rely on solving the energy balance at the glacier surface. This requires extensive data from field surveys and meteorological stations. An alternative is temperature index modelling, where temperature is considered as a proxy for all melt processes. However, the approximations developed are not necessarily representative of Himalayan glacier mass balance processes. With the increasing data from direct glaciological measurements, geodetic methods, and indirect remote sensing techniques, there has been an increase in statistical and machine learning solutions. While it is a promising alternative, it remains underutilised due to challenges such as regionally varying limitations in ground-truth data and the perceived black-box nature of machine learning models. This thesis contributes to the evolving landscape of ML modelling of glacier mass balance through a set of objectives, described as follows: through the first objective, the thesis establishes techniques for machine learning-based glacier mass balance estimation by systematically comparing model performance under varying data availability scenarios. It demonstrates the advantages of non-linear ensemble tree-based models over traditional linear regression. This aligns with the results from recent statistical approaches to glacier mass balance modelling. The best-performing gradient boosted regression model resulted in a testing RMSE value of 1.071m.w.e, MAE values of 0.774m.w.e, and an R2 value of 0.71. The study also highlights the role of transfer learning applied with neural networks in extending knowledge from data-rich to data-sparse regions. Through the second objective, the thesis extends the indirect remote sensing methods to spatial scales by incorporating non-linear interactions of Equilibrium Line Altitude and Accumulation Area Ratio with glacier mass balance. In addition to Equilibrium Line Altitude and Accumulation Area Ratio, additional predictors such as glacier area and aspect were also included. We conclude that the models augmented using gradient boosted regression outperform existing models at all spatial scales. Further, we demonstrate the improved performance of the model at glacier level and higher-order spatial basin scales, especially when forced using additional inputs of area and aspect (with a best performance RMSE of 0.909 m w.e). Finally, as the third objective, a feature importance study of various meteorological variables was conducted over the Himalayas. We observed that, beyond temperature, wind speed, specific humidity, and shortwave radiation were also important drivers of glacier mass balance in the Himalayas. These results highlighted the need to incorporate additional variables into the classical temperature index model. Thus, a novel model - XAITI - was developed, integrating the spatially distributed set of meteorological drivers of glacier mass balance into an enhanced temperature index modelling framework. By addressing primary limitations in data quality, spatial generalizability, and model transparency, this work bridges the gap between process-based and data-driven approaches. It introduces scalable methodologies that can be extended to other glaciated regions of the world.