Rainfall-Runoff Modeling in a Himalayan Watershed Using Adaptive Neuro-Fuzzy Inference System with Gamma Test-based Input Selection

In this study, rainfall-runoff modeling was done using adaptive neuro-fuzzy inference system (ANFIS) in a Himalayan watershed of Ramganga catchment in Uttarakhand, India. The daily observed rainfall and runoff data of monsoon season for 10 years (June 2000 to September 2009) were used for training and testing of the ANFIS models. Gamma test (GT) was applied for selection of the best combination of input variables prior to modeling the rainfall-runoff process. The performance of the developed ANFIS models was evaluated by comparing the simulated and observed runoff values by using statistical indices, like correlation coefficient (r), root mean square error (RMSE), Nash-Sutcliffe efficiency (NSE) and pooled average relative error (PARE). Of the total 14 models, developed using seven membership functions, model ANFIS10 was found to be the best performing with gauss-3 membership function. For the best model, values of r, RMSE, NSE and R² were found to be 0.97, 1.03 mm, 0.92 and 0.94, respectively, during calibration stage. Similarly, values of r, RMSE, NSE and R² for the best model during testing stage were found to be 0.95, 0.70 mm, 0.89 and 0.90, respectively. The PARE value of 0.024% and 0.005% for best model during testing and training phase, respectively, indicated negligible (0.024% and 0.005%) over-prediction highlighting its accuracy and reliability. It was found that the runoff of the present day depends on current day rainfall as well as previous two consecutive days' rainfall and runoff. Comparison of the results of this study with that reported in earlier study revealed a higher accuracy of ANFIS model than that of multilayer perceptron based neural network, radial basis function based neural network and multiple linear regression. This study concluded that ANFIS model can be effectively used for rainfall-runoff modelling with good accuracy in the study watershed.