Interpretable Deep Learning for Glacier Mass Balance: Temporal Attention Patterns in Central Asia

Abstract. Glaciers in Central Asia are critical water resources, and their retreat is closely linked to glacier-related hazards such as debris flows, glacial outburst floods, and landslides, yet their response to climate change is poorly understood due to complex temporal dependencies. Although the accuracy of mass-balance estimation has increased recently with advances in machine learning, current methods provide only a limited understanding of when changes in mass balance are driven by climate variables. This knowledge is crucial for hazard assessment and adaptation planning. The current study introduces Temporal Fusion Transformers (TFT V2), a deep learning architecture with interpretable attention mechanisms, to predict mass balance across 43,018 glaciers in seven Central Asian mountain ranges using 2000–2018 climate reanalysis and geodetic mass balance data. In this study, the TFT V2 model achieves R² = 0.73 (RMSE = 0.21 m w.e., MAE = 0.11 m w.e.) on independent test data while providing reliable uncertainty quantification (calibration score = 0.94, coverage within±2–3 % of nominal levels). Critically, attention weights reveal that summer months (June–August) contribute 35 % of predictive signal— 3× more than winter months (18 %)—and identify spring melt onset (April–May, 28 % importance) as critical for annual balance. In Central Asian mountain regions, where hazard risk is driven by greater precipitation and earlier melt onset, these temporal patterns are directly linked to observed increases in mudflows and landslides throughout the spring. Due to variations in hazard vulnerability, regional analysis shows spatial variety in temporal patterns, with the Tian Shan showing larger summer concentrations than the Pamir. The main contribution of the model is the identification of the temporal cascade of glacier mass-balance factors, despite achieving competitive predictive performance (R2=0.73). For the first time, we quantify the dominant influence of spring melt onset (April–May) on peak summer ablation, a dynamic link previously hypothesised but not demonstrated at a regional scale. This indicates that deep learning can balance competitive predictive performance with temporal insights and uncertainty awareness unavailable from traditional ML approaches.