Identifying dominant parameters in SWAT across subbasin and HRU scales using a two-step deep learning-assisted spatial sensitivity analysis

Distributed hydrological models require sensitivity analyses that explicitly account for spatial heterogeneity, yet such analyses are often constrained by high computational demands. This study presents a two-step, deep learning-assisted spatial sensitivity analysis (SSA) framework to identify dominant parameters across multiple spatial scales. Using the Soil and Water Assessment Tool (SWAT) in the Jinghe River Basin as a case study, the Morris method was first applied with a spatially lumped strategy to screen influential parameters. Subsequently, SSA was conducted using the Sobol' method with multilayer perceptron (MLP) surrogates to evaluate parameter sensitivities under both subbasin- and hydrologic response unit (HRU)-scale parameterizations. The surrogate models emulated SWAT with high accuracy for 195 subbasin-scale and 2559 HRU-scale parameters, enabling efficient estimation of Sobol' sensitivity indices. Results reveal pronounced spatial heterogeneity in parameter sensitivities. Sensitivity hotspots are consistently concentrated in gauge-proximal subbasins, while HRU-scale analysis further resolves localized controls. Robustness analyses demonstrate that the identified sensitivity patterns remain stable across different performance metrics, NSE-constrained posterior parameter distributions, alternative observation configurations, and varying numbers of screened parameters. These findings confirm the reliability of the proposed two-step framework and highlight its capability to efficiently diagnose dominant hydrological controls in high-dimensional distributed models. The methodology provides a transferable and computationally feasible approach for sensitivity assessment, supporting scale-aware calibration and improving the interpretability and predictive reliability of distributed hydrological simulations.