Identifying Dominant Parameters Across Space and Time at Multiple Scales in a Distributed Model Using a Two-Step Deep Learning-Assisted Time-Varying Spatial Sensitivity Analysis

Abstract. Distributed models require parameter sensitivity analyses that capture both spatial heterogeneity and temporal variability, yet most existing approaches collapse one of these dimensions. We present a two-step, deep learning-assisted, time-varying spatial sensitivity analysis (SSA) that identifies dominant parameters across space and time. Using SWAT for runoff simulation of the Jinghe River Basin, we first apply the Morris method with a spatially lumped strategy to screen influential parameters and then perform SSA using a deep learning-assisted Sobol' method for quantitative evaluation. A key innovation lies in the systematic sensitivity evaluation with parameters represented and analysed at both subbasin and hydrologic response unit (HRU) scales, enabling explicit treatment of distributed parameters at their native spatial resolutions. To reduce computational burden, two multilayer perceptron surrogates were trained for 195 subbasin and 2,559 HRU parameters, respectively, allowing efficient time-varying SSA of NSE-based Sobol' indices over 3- and 24-month rolling windows during 1971–1986. Results reveal structured, scale-dependent controls: spatially, sensitivity hotspots are coherent between scales but become more localized at the HRU level, reflecting heterogeneity in land use, soils, and topography; temporally, sensitivities fluctuate with runoff in the 3-month window, while event-scale variations are smoothed in the 24-month window, yielding more persistent patterns governed by storage and routing processes. The proposed framework provides a computationally efficient and unified approach for identifying scale-dependent sensitivity hotspots and hot moments, thereby supporting targeted calibration and enhancing the interpretability and predictive robustness of distributed models under nonstationary conditions.