Comparison of machine learning and process-based SWAT model in simulating streamflow in the Upper Indus Basin

This study appraised and compared the performance of process-based hydrological SWAT (soil and water assessment tool) with a machine learning-based multi-layer perceptron (MLP) models for simulating streamflow in the Upper Indus Basin. The study period ranges from 1998 to 2013, where SWAT and MLP models were calibrated/trained and validated/tested for multiple sites during 1998-2005 and 2006-2013, respectively. The performance of both models was evaluated using nash-sutcliffe efficiency (NSE), coefficient of determination (R-2), Percent BIAS (PBIAS), and mean absolute percentage error (MAPE). Results illustrated the relatively poor performance of the SWAT model as compared with the MLP model. NSE, PBIAS, R-2, and MAPE for SWAT (MLP) models during calibration ranged from the minimum of 0.81 (0.90), 3.49 (0.02), 0.80 (0.25) and 7.61 (0.01) to the maximum of 0.86 (0.99), 9.84 (0.12), 0.87 (0.99), and 15.71 (0.267), respectively. The poor performance of SWAT compared with MLP might be influenced by several factors, including the selection of sensitive parameters, selection of snow specific sensitive parameters that might not represent actual snow conditions, potential limitations of the SCS-CN method used to simulate streamflow, and lack of SWAT ability to capture the hydropeaking in Indus River sub-basins (at Shatial bridge and Bisham Qila). Based on the robust performance of the MLP model, the current study recommends to develop and assess machine learning models and merging the SWAT model with machine learning models.