Evaluation of neuro-fuzzy and Bayesian techniques in estimating suspended sediment loads

Sediment is considered the largest surface water pollutant by volume, which is crucial for surface water planning and management. Different management scenario evaluations require multiple in-stream suspended sediment forecasts and estimations. Physically-based models are considered to be good modeling techniques for suspended sediment estimation; nevertheless, they require a large number of parameters and intensive calculations. This study aims to enhance suspended sediment predicting techniques using efficient fusion modeling that can be used for evaluations by watershed managers and stakeholders. Adaptive neuro-fuzzy inference system (ANFIS) and Bayesian regression models were tested to find the best alternative to a calibrated and validated Soil and Water Assessment Tool (SWAT) model to predict suspended sediment loads in the Saginaw River watershed. For both methods, four different method-types were tested, namely General, Temporal, Spatial and Spatiotemporal. Results of the study showed that both methods can be used as good alternatives to the SWAT model at the global level for watershed estimations. The best suspended sediment replicating models, the Bayesian Spatiotemporal and ANFIS Spatial, produced results with Nash-Sutcliffe model efficiency values of 0.95 and 0.94, respectively. For the subbasin level, Bayesian and ANFIS techniques showed satisfactory results for 84 and 77 subbasins, respectively, out of 155 subbasins in the watershed. Box-Cox transformation of suspended sediment load values, made the use of the Bayesian model feasible and improved the prediction of the ANFIS models. However, suspended sediment data exhibited a bimodal distribution after transformation, making the modeling process challenging and complex.