Remote sensing of lake surface water temperature reveals divergent hydroclimatic associations with lake ice phenology across China

Lake ice phenology is a sensitive indicator of climate change, yet its regional hydroclimatic associations remain poorly understood. Here, we reconstructed daily lake surface water temperature (LSWT) from MODIS for 583 lakes across China and derived freeze-up date (FUD), break-up date (BUD), and ice cover duration (ICD) for 376 seasonally ice-covered lakes. Robustness was evaluated with sensitivity tests, meteorological observations, transition-period comparisons, and passive-microwave ice records; Mann-Kendall tests, Theil-Sen slopes, VIF-constrained XGBoost-SHAP, and structural equation modeling assessed trends and regional associations. FUD was delayed in all three ice-covered regions, whereas BUD was delayed in the Tibetan Plateau Lake Region (TPR) but advanced in the Mongolian-Xinjiang Lake Region (MXR) and Northeast Lake Region (NLR), resulting in shorter ICD in all regions. Regional mean ICD declined by 0.03, 0.35, and 0.47 d yr -1 in the TPR, MXR, and NLR, respectively; significant lake-level changes were concentrated in the TPR, with 72 and 33 lakes showing significant FUD and BUD trends. All nine models retained positive hold-out R 2 (0.535–0.960). SHAP attributions differed by region: radiative variables dominated in the TPR, the MXR showed metric-specific associations with wind, snowfall, geographic, and thermal , and geographic, radiative, and morphometric variables dominated in the NLR. Structural equation models indicated stronger precipitation-related associations in the NLR and snowfall-related associations in the MXR and TPR. These results reveal pronounced regional heterogeneity in lake ice responses to warming and the value of integrating remote sensing, interpretable machine learning, and pathway modeling to resolve these contrasts.