Insight into glacio-hydrologicalprocesses using explainable machine-learning (XAI) models

Abstract

The glacio-hydrological process is essential in the global water cycle but is complex and poorly understood. In this study, we couple the deep Shapley additive explanation (SHAP) with a long short-term memory (LSTM) model to construct a machine-learning (XAI) framework that describes the glacio-hydrological process in Urumqi Glacier No. 1, China. The XAI framework reveals 1) the dominant hydro-meteorological factors have a five-month lead time, and each factor has its own active time and degree of contribution; 2) the temperature and precipitation within the lead time dominate the process; 3) identifiable combination of the factors, instead of extreme events themselves, creates the extreme glacio-hydrological phenomena. Generally, the glacial meltwater replenishes the glacial stream runoff, which is influenced by many environmental factors. In particular, the runoff responds to the change in the glacier mass balance with hysteresis within five months. Overall, the temperature and precipitation within the lead time (4–5 months) dominate the runoff processes. This study quantifies the Contribution of each input in the glacio-hydrological process and provides valuable insight into the interaction of various hydro-meteorological factors.