No abstract

Extreme weather conditions like floods and droughts call for careful planning and management of water resources in order to prevent fatalities and other negative effects. Modern soft computing and machine learning approaches have provided a solution for simulating these hydrological phenomena despite the complexity and non-linear character of these phenomena, which depend on diverse parameters. Distributed or semi-distributed models for large-size watershed areas with geographical irregularity and heterogeneity necessitate a substantial amount of high-quality spatial data. This research uses 40 years (1981–2020) of daily rainfall-runoff data to illustrate the application of two data-driven models, random forest regression (RFR) and feed-forward neural network (FFNN), for semi-arid, large-size watershed areas. To understand the effect of input data, different input–output combinations were considered to simulate eight rainfall-runoff models. Results show that both RFR and FFNN models have successfully performed but RFR model performance is best with correlation coefficient values of 0.9928 (M6) and 0.9926 (M1).