Comparison of a data-based model and a soil erosion model coupled with multiple linear regression for the prediction of reservoir sedimentation in a semi-arid environment

Abstract

Reservoir sedimentation is a crucial challenge in planning and managing sustainable surface water resources in arid and semi-arid regions and must be assessed with accuracy. Both data-based models and conceptual models can be valuable tools for predicting reservoir sedimentation. In this study, we used an artificial neural network (ANN) approach and a modified Universal Soil Loss Equation coupled with multiple linear regression (MUSLE-MLR) model to predict yearly sedimentation in the Sidi Mohammed Ben Abdellah reservoir, located in a semi-arid region of Morocco. To construct the MUSLE-MLR model, we first calibrated and validated the MUSLE on 32 storms at four gauging stations upstream of the dam to estimate the sediment yield at these four gauging stations; we then developed the MLR model for combining sediment yield and reservoir sedimentation. The results of this model were then compared with the performance of the ANN model that was trained and validated over the periods 1975–2008 and 2009–2015, respectively. The comparison revealed that the calibrated MUSLE model is fairly useful to predict sediment yield at the watershed level. However, comparison of the two models during the validation process showed that the ANN (R2 0.91, Nash–Sutcliffe Efficiency [NSE] 0.820) is more accurate and more suitable than the MUSLE-MLR model (R2 0.819, NSE − 1.592) to predict reservoir sediment in the Sidi Mohammed Ben Abdellah reservoir. The findings of this study contribute to the armamentarium of potential tools that can be used to predict and manage reservoir sedimentation at the watershed and reservoir levels in a semi-arid context.