Spatiotemporal Reconstruction of Annual Glacier Mass Balance in Central Asia (2000–2020) Using Machine Learning

Glaciers in High-mountain Asia play a critical role in both climate change studies and regional water resource management. However, detailed observations over a large spatial extent remain scarce. In this study, we reconstruct annual glacier-wide mass balance from 2000 to 2020 for glaciers larger than 0.1 km² across the Tien Shan and Pamir using machine learning (ML) techniques. Five ensemble ML and a deep neural network models were tested, with XGBoost demonstrating the best performance and thus selected for the reconstruction of the glacier mass balance time series. Predictor variables included meteorological data from the ERA5-Land data set and topographic features. The results indicate an average mass loss of −0.39 m water equivalent (m w.e.) per year for the studied period, with the highest losses observed in the Djungar Alatau (−0.68 m w.e. yr⁻¹), and the lowest in the eastern Pamir (−0.10 m w.e. yr⁻¹). Additionally, the results reveal that small glaciers (area <1 km²) experience more rapid mass loss. The temporal evolution of glacier mass balance exhibits, on average, an acceleration but with spatiotemporal variability. Variable importance analysis identified glacier elevation and geographic location as the dominant factors influencing mass balance, followed by the temperatures of July and August. This work further advances the application of ML methods in glaciology, enhancing our understanding of regional glacier mass balance.