Explainable deep learning for predicting lake expansion in the Siling Co Endorheic basin: Unraveling driver shifts and hydro-climatic tipping points

Study region The Siling Co Endorheic Basin (SCEB) on the central Tibetan Plateau is a critical terminal closed-flow catchment within the Asian Water Tower. Highly sensitive to hydro-climatic water balances, the basin has undergone rapid, cryosphere-fed lake expansion. This expansion coincides with significant regional warming and wetting. Study focus To clarify alpine lake evolution mechanisms and overcome the black-box limitations and phase lags of traditional models, this study forecasts the monthly total lake area sequence of the SCEB (2000–2021). We developed the CAF(PINN) framework, a physics-informed approach that embeds the water balance principle over a hybrid architecture consisting of a 1D-CNN, an attention mechanism, and a frequency-domain multi-layer perceptron. Additionally, we quantified interrelationships within lacustrine hydrological processes and physical tipping points using PLS-SEM and Transformed Generalized Additive Models (TGAM). New hydrological insights The CAF(PINN) framework achieved a high Nash–Sutcliffe Efficiency of 0.91, outperforming benchmark models while preserving physically consistent behavior. The model identified a distinct hydrological transition in 2013, suggesting a regime shift toward a meltwater-dominated system. Prior to 2013, lake variations were largely associated with precipitation fluctuations while the cryosphere buffered thermal inputs. Post-2013, however, rising temperatures were closely linked to increased glacier melt and permafrost thaw as primary contributors to lake recharge. This transition suggests an intensifying, non-linear coupling between cryospheric change and regional hydrology, highlighting the expanding influence of thermal factors on water cycles within the Asian Water Tower.