Improving Driver Attribution for Water-Level Variations in Qinghai Lake Through Hydrological Process-Period Analysis

Daily water-level changes in alpine closed lakes integrate lake-surface exchange, cryosphere processes, basin storage, and runoff routing, yet fixed calendar seasons and synchronous predictors may mix contrasting hydrological states and misalign antecedent forcing. This study developed a Hydro-Lag framework for Qinghai Lake, China’s largest inland lake, during 2014–2023 to improve daily reconstruction and driver attribution through dynamic hydrological process periods and time-lag effects. Satellite altimetry, optical and thermal imagery, hydrometeorological reanalysis, lake ice, and runoff data were integrated with water-balance constraints. The frozen (T1) and ablation (T2) periods were identified from lake ice transitions and a positive degree-day (PDD) threshold estimated from ice thickness. Meanwhile, basin wetness, represented by the antecedent precipitation index (API), distinguished the wet (T3) and dry (T4) periods. Candidate lags from cross-correlation analysis were refined through repeated cross-validation and used in an ensemble of extreme gradient boosting (XGBoost) and ExtraTrees; cross-fitted predictions were interpreted with SHapley Additive exPlanations (SHAP). Alignments differed markedly: selected core variables in T1 aligned mainly at 10–12 days; T2 used 1–2 days runoff, soil-moisture, and evapotranspiration offsets together with a same-day 14-day M_snow accumulation; T3 runoff and soil moisture at 3–5 days and precipitation-related variables at 8–9 days; and T4 antecedent basin variables at 11–24 days. Against an independently tuned SeasonRAW baseline, Hydro-Lag increased test R2 from 0.551 to 0.793 and reduced RMSE from 14.26 to 8.51 mm d−1. OOF SHAP indicated period-dependent patterns in modeled predictor contributions: mixed contributions in T1, wetness–storage emphasis in T2–T3, and antecedent storage and thermo-evaporative contributions in T4. Conditional responses showed thresholds and sign reversals across thermal, wetness, and event-magnitude gradients. The framework provides a process-aware basis for assessing daily lake–basin interactions in data-sparse alpine closed basins.