Shifted dominant flood drivers of an alpine glacierized catchment in the Tianshan region revealed through interpretable deep learning

Abstract The Kumalak River, a typical alpine glacierized catchment in the Tianshan region, experiences complex flooding driven by glacier meltwater, snowmelt, and rainfall. However, the mechanisms driving these processes under climate change remain unclear. To address this, a SWAT-Glacier hydrological model and a degree–day factor model were used for snowmelt, glacier meltwater, and rainfall calculations. Two Long Short-Term Memory (LSTM) models (LSTM-SG and LSTM-DDF) were developed using these inputs, and additive decomposition and integrated gradient methods were applied to interpret flood mechanisms. Glacier meltwater was found to dominate annual maximum flood (AMF) events, while snowmelt drove annual spring maximum flood (AMFSp) events. For AMF events (1960–2018), contributions were 10.01–12.21% from snowmelt, 60.49–60.92% from glacier meltwater, and 26.86–29.50% from rainfall. For AMFSp events (1961–2018), contributions were 48.49–56.08% from snowmelt, 16.12–22.08% from glacier meltwater, and 27.79–29.42% from rainfall. These findings provide critical insights for enhancing flood prediction and optimizing water resource management.