Abstract With the growing availability of global glacier thickness data sets, there is a need for fast and adaptable approach to assess how glacier thickness changes over time. This study demonstrates how machine learning (ML) models can be trained using existing thickness data sets and later applied with updated inputs to quantify the changes in glacier thickness. This study introduces an innovative skeletonization distance approach, which converts glacier outlines into thin central structure and computes each pixel's distance from this structure to represent glacier geometry more effectively. This new input variable, combined with elevation and slope, was used to train four tree‐based ML models: Random Forest, XGBoost (XGB), AdaBoost‐RF, and AdaBoost‐DT. Among them, XGB performed best, achieving R 2 values of 0.82–0.88 (training) and 0.73–0.82 (testing). The best‐performing XGB model was then used to analyze glacier thickness changes between 2000 and 2015 for the Bara Shigri, Gangotri, and Zemu glaciers. The analysis revealed heterogeneous thinning concentrated in debris‐covered ablation zones of the study glaciers. Incorporating skeletonization improved test accuracy by up to 13%. To test transferability, the approach was applied to Aletsch (Switzerland), Koxkar (China), and Saskatchewan (Canada) glaciers, effectively capturing regional thinning patterns consistent with previous studies. The trained models can be directly applied with Digital Elevation Models and glacier boundaries of different years, allowing researchers to estimate glacier thickness changes.
Singh et al. (2026) studied this question.