Basin runoff calibration using gradient boosting decision trees: Impact of dual gridded potential evapotranspiration sources on model performance in Iznik Lake Basin, Türkiye

Study region This study focuses on the Iznik Basin, an endorheic catchment in a semi-humid region of Bursa, Türkiye, characterized by significant anthropogenic influence and hydrological sensitivity. Study focus This research evaluates the impact of gridded potential evapotranspiration (GPET) products, specifically the reanalysis-based ERA5-Land dataset with the satellite-based GLEAM on daily streamflow prediction. Three gradient boosting algorithms (CatBoost, XGBoost, and LightGBM) were implemented. Hyperparameters were optimized via Optuna utilizing temporal cross-validation to maximize performance across multiple metrics: Nash–Sutcliffe efficiency (NSE), LogNSE, Kling–Gupta efficiency (KGE), and percent bias (PBIAS). To ensure model transparency, SHapley additive exPlanations (SHAP) were employed to interpret feature importance, while Jacobian-based sensitivity analysis quantified the sensitivity of runoff predictions to GPET inputs. New hydrological insights LightGBM exhibited superior performance in simulating daily streamflow, achieving a NSE of 0.69 and a KGE of 0.77, though CatBoost surpassed both LightGBM and XGBoost in predicting low-flow dynamics. The integration of GLEAM inputs resulted in an approximately 30% reduction in PBIAS propagation compared to ERA5-Land, which demonstrated a systematic negative deviation. SHAP identified GPET as the third most influential predictor, primarily affecting model behavior during recession periods. Jacobian analysis revealed a negative correlation between GPET and streamflow. These findings underscore that the choice between reanalysis-based and satellite-based GPET is a critical determinant in the uncertainty budget of data-driven hydrological frameworks, particularly in data-sparse regions.