Comparison of soil and water assessment tool (SWAT) and multilayer perceptron (MLP) artificial neural network for predicting sediment yield in the Nagwa agricultural watershed in Jharkhand, India

Abstract

► Performance evaluation and comparison of the process based model SWAT and the data driven model MLP have been carried out for Nagwa watershed in India for simulating monthly sediment yield . ► The annual average measured sediment yield was 3.7 t/ha. The annual average simulated sediment yield was 3.1 and 5.0 t/ha for MLP and SWAT model, respectively. ► Three scenarios of the input set of data were considered and it was observed that daily rainfall and average daily discharge together improved the performance of the MLP model in the estimation of sediment load. Inclusion of previous day discharge did not improve the result. ► For simulating monthly sediment yield at a single outlet of a watershed, the MLP model performed better than SWAT, especially when simulating lower values of sediment yield. The present study was conducted in the Nagwa watershed in Jharkhand state, India. The watershed has been identified as a sensitive area for sediment and non-point source pollution. Damodar Valley Corporation (DVC), Hazaribagh, India has taken initiatives to implement several soil and water conservation measures. A calibrated and validated model to simulate hydrological processes will be a great help to the concerned watershed managers. The objectives of this study were to compare the monthly sediment yield simulation results from the soil and water assessment tool (SWAT) and the multilayer perceptron (MLP) artificial neural network model during the calibration (1993–2004) and validation periods (2005–2007), and determine the most appropriate model for the watershed. The annual average measured sediment yield was 3.7 t/ha. The annual average simulated sediment yield was 3.1 and 5.0 t/ha for MLP and SWAT model, respectively. Both models generally provided good correlation and model efficiency for simulating monthly sediment yield during calibration and validation. For the SWAT model the coefficient of determination ( R 2 ) and Nash-Sutcliffe simulation efficiency ( NSE ) values were 0.78 and 0.76 during calibration and 0.68 and 0.66 during validation, respectively. The MLP model performed better than SWAT with R 2 and NSE values of 0.84 and 0.76 during training and 0.77 and 0.74 during validation periods, respectively. In the present study, the MLP artificial neural network model was a better model than SWAT for simulating sediment yield from the single outlet of this watershed based on calibration and validation results. The water resource managers and different stakeholders can use this validated model for planning and implementing appropriate soil and water conservation measures.