Streamflow modeling of a large arid catchment using semi-distributed hydrological model and modular neural network

Abstract

Calibration and validation of hydrological models for simulating stream flow can usually be a promising procedure for future sustainable watershed development. Therefore, development of hydrological models with attributed capabilities is vital to explore the models. Recently, arid climate regions are facing critical water resource problems due to elevated water scarcity. The main objective of this research is to compare the Soil and Water Assessment Tool (SWAT), a knowledge driven by semi-distributed hydrological model, with the Modular Neural Network (MNN), a data driven technique, in predicting the daily flow in arid and large scale. Development of SWAT required digital elevation map, hydro-meteorological data, land use map, and soil maps; whilst, the MNN only needed hydro-meteorological data. For both models, a sensitivity analysis that included both calibration and validation with individual uncertainty evaluation methods was carried out. Generally, results for relative errors such as Nash-Sutcliffe, coefficient of determination and percent of bias favored the SWAT for the validation period. Not only that, the absolute error criteria such as root mean square error, mean square error and mean relative error obtained were close to zero for the SWAT as well within the same period. The mean absolute error for both models was similar during the validation period. Results of the uncertainty evaluation were in satisfactory range. Both models had given similar trend for flow prediction during the validation period. Results of box plot, according to 50% (median) of daily flow, showed that both models had respectively overestimated (MNN) and underestimated (SWAT) the daily flow during validation period. Evaluation on runoff volume for each year showed that both models had a one-year underestimation and three-year overestimation in the same period. However, the overestimation of MNN was more obvious. As a conclusion, this study showed that both models have promising prediction performance for daily flow in a large scale watershed with arid climate