A graph convolutional long short-term memory (GC-LSTM) surrogate model was developed to emulate daily flow generated by the Soil and Water Assessment Tool (SWAT) for the East Fork Shoal Creek (EFS) watershed. The EFS SWAT model represents an 80.22 km² agricultural watershed comprising 23 subbasins and 678 hydrologic response units (HRUs), with row crops covering 78.1% of the watershed area. A Sobol low-discrepancy sampling design was used to generate 1,024 realizations across 15 influential flow-related parameters, providing balanced coverage of the multidimensional parameter space. Each realization produced 2,191 daily flow values spanning 2014-2019.
For each prediction day, the surrogate model used meteorological forcing and watershed conditions from the preceding 14 days to capture short-term hydrologic memory associated with antecedent precipitation, soil moisture, groundwater storage, and channel-routing effects. Inputs included daily precipitation and temperature, 15 SWAT parameters representing key hydrologic processes, and spatial attributes describing subasins' characteristics. Graph convolution aggregated information from each subbasin and its upstream contributors to generate subbasin-scale representations of hydrologic conditions. These representations were processed by a 32-unit LSTM network to predict daily flow simultaneously at the outlets of all 23 subbasins.
The dataset was divided into training (70%), validation (15%), and independent test (15%) subsets. Learning-curve analysis identified 512 realizations as sufficient for training the pre-calibration surrogate model. On the independent test set, the model achieved macro Nash-Sutcliffe efficiency (NSE) and Kling-Gupta efficiency (KGE) values of 0.66 and 0.65, respectively, where macro metrics represent unweighted averages across all subbasin outlets. At the watershed outlet, NSE and KGE were 0.67 and 0.68, respectively, with a percent bias of 1.9%. Future work will involve calibrating the surrogate model against observed daily flow and transferring calibrated parameter estimates and associated uncertainty bounds to the EFS SWAT model.