Effects of soil data resolution on agricultural hydrological simulations in a groundwater-overexploited region: A SWAT and XGBoost-SHAP analysis

Soil is the critical interface governing evapotranspiration, groundwater recharge, and crop water uptake in terrestrial hydrological systems. The effects of soil data resolution on simulation accuracy, however, are highly process- and region-specific, and systematic evaluations jointly addressing horizontal and vertical dimensions across multiple targets remain limited. This study tested three soil configurations varying in horizontal resolution (10 km and 1 km) and vertical discretization (two and five layers) within a modified Soil and Water Assessment Tool (SWAT) model in a groundwater-overexploited piedmont plain of the North China Plain, evaluating performance against evapotranspiration (ET a ), groundwater table depth, and crop yield. An XGBoost-SHAP surrogate modeling framework was applied to provide parameter-level attribution of resolution effects on SWAT simulation outputs. Groundwater depth achieved the highest accuracy across all configurations, with an NSE of 0.91 and an R 2 of 0.91–0.92, reflecting physically constrained subsurface dynamics governed primarily by hydrogeological parameters. ET a showed intermediate performance ( R 2 > 0.75) but NSE ranging from 0.48 to 0.55, where soil-atmosphere coupling introduced additional uncertainty. Crop yield exhibited the greatest variability (normalized root mean square error 18.8–23.1%), where biological processes amplified sensitivity beyond purely physical controls. The 1 km two-layer dataset gave the best ET a accuracy a nd the most balanced performance among the three configurations. SHAP attribution revealed that horizontal resolution improvement systematically elevated soil feature importance relative to climate forcing for ET a and yield, while groundwater dynamics remained dominated by hydrogeological parameters regardless of soil configuration. Ablation analysis established horizontal resolution as the dominant accuracy dimension. Refining horizontal resolution from 10 km to 1 km improved surrogate accuracy for all three targets, with gains in R² of 0.043 for ET a , 0.002 for groundwater, and 0.083 for yield, whereas adding a fifth vertical layer reduced accuracy for every target, by 0.048, 0.008, and 0.057 respectively, introducing redundant information without predictive benefit. By resolving how resolution effects attenuate along the Crop-Soil-Groundwater continuum, this study provides a transferable basis for prioritizing horizontal over vertical refinement in soil dataset selection for agricultural hydrological modeling.