Runoff and Sediment Yield Processes in a Tropical Eastern Indian River Basin: A Multiple Machine Learning Approach

Abstract

Tropical Indian river basins are well-known for high and low discharges with high peaks of flood during the summer and the rest of the year, respectively. A high intensity of rainfall due to cyclonic and monsoon winds have caused the tropical Indian rivers to witness more runoff. These rivers are also known for carrying a significant amount of sediment load. The complex and non-linear nature of the sediment yield and runoff processes and the variability of these processes depend on precipitation patterns and river basin characteristics. There are a number of other elements that make it difficult to forecast with great precision. The present study attempts to model rainfall–runoff–sediment yield with the help of five machine learning (ML) algorithms—support vector regression (SVR), artificial neural network (ANN) with Elman network, artificial neural network with multilayer perceptron network, adaptive neuro-fuzzy inference system (ANFIS), and local linear regression, which are useful in river basins with scarce hydrological data. Daily, weekly, and monthly runoff and sediment yield (SY) time series of Vamsadhara river basin, India for a period from 1 June to 31 October for the years 1984 to 1995 were simulated using models based on these multiple machine learning algorithms. Simulated results were tested and compared by means of three evaluation criteria, namely Pearson correlation coefficient, Nash–Sutcliffe efficiency, and the difference of slope. The results suggested that daily and weekly predictions of runoff based on all the models can be successfully employed together with precipitation observations to predict future sediment yield in the study basin. The models prepared in the present study can be helpful in providing essential insight to the erosion–deposition dynamics of the river basin.