On the complexities of sediment load modeling using integrative machine learning: Application of the great river of Loiza in Puerto Rico

Sediment transportation in water bodies may cause many problems for the water resources projects and damage the environment. Hence, modeling sediment load components, including suspended sediment load (SSL) and bedload (BL) in rivers is of prime importance. Effective modeling of SSL and BL remains a challenging task due to their complex hydrological process. On this account, this study aims to appraise the potential of conventional machine learning (ML) models including adaptive neuro-fuzzy inference system (ANFIS), support vector regression (SVR), and their integrative version with nature optimization algorithm called genetic algorithm (GAANFIS and GA-SVR) for SSL and BL prediction. Two traditional models are developed for modeling verification including the sediment rating curve (SRC) and multiple linear regression (MLR). The modeling results are assessed using four statistical measures (e.g., root mean square error (RMSE), mean absolute error (MAE), Nash-Sutcliffe Efficiency (NSE), and coefficient of determination (R-2)), diagnostic analysis (scatter plots and Taylor diagram), and evaluation of the dependence of the state of the river flow-sediment system (hysteresis analysis). Based on the attained predictability performance, the integrative ML models reveal a superior prediction capacity in comparison with the standalone ANFIS, SVR, and the traditional models. In quantitative evaluation, the proposed integrative ML models indicate a remarkable prediction enhancement approximately 44% mean magnitude based on the MAE metric against the SRC traditional model for both the SSL and BL predicted values. Overall, the current investigation evidences the potential of the nature-inspired algorithm as a hyper-parameter optimizer for ML models that produce a reliable and robust predictive model for sediment concentration quantification.