Document Type : Research Paper
Author
Assistant Professor, Soil Conservation and Watershed Management Research Department, Kermanshah Agricultural and Natural Resources Research and Education Center, Agricultural Research, Education and Extension Organization (AREEO), Kermanshah, Iran
Abstract
Introduction
Suspended sediment transport constitutes a critical challenge in water resources engineering, causing reservoir capacity reduction, water quality degradation, and hydraulic infrastructure damage. Conventional sediment rating curve (SRC) methods rely on simplifying assumptions that frequently fail to capture nonlinear and transient flow-sediment dynamics. Recent advances in data-driven approaches, particularly artificial neural networks (ANNs) and adaptive neuro-fuzzy inference systems, have demonstrated substantial potential for modeling such complex hydrological processes. This study develops and rigorously evaluates a hybrid framework that integrates empirical SRCs with a Coactive Neuro-Fuzzy Inference System (CANFIS) to enhance predictive accuracy and model robustness at the Glinak hydrometric station in the Taleghan watershed.
Materials and methods
The analysis utilized 800 paired observations of streamflow and suspended sediment discharge collected from Glinak station between 1971 and 2023. Data preprocessing involved outlier removal, homogeneity assessment using the standard normal homogeneity test, and normalization. Six empirical SRC methods were calibrated and evaluated: logarithmic-linear regression, multi-segment regression, Logged Mean Load within Discharge Classes Method (LMLWDC), modified FAO correction factor, parametric correction factor, and non-parametric correction factor. Performance evaluation employed the coefficient of determination (R²), Nash-Sutcliffe efficiency (NSE), and relative mean error (RME). In parallel, six ANN architectures—Multilayer Perceptron (MLP), General Feed-Forward (GFF), Radial Basis Function (RBF), Support Vector Machine (SVM), Self-Organizing Feature Map (SOFM), and CANFIS—were trained using a 75%-25% train-test split. Model assessment utilized normalized mean squared error (NMSE), mean absolute error (MAE), and Pearson's correlation coefficient (R). Two hybrid configurations were examined: Hybrid Model 1 used the optimal SRC output as the sole input to CANFIS, whereas Hybrid Model 2 incorporated both SRC output and streamflow as inputs.
Results and discussion
A detailed interrogation of the findings at the Glinak hydrometric station reveals that conventional hydrological models encounter substantive structural constraints when estimating suspended sediment load under conditions of extreme data dispersion, as evidenced by a coefficient of variation of 206.75%. Pronounced fluctuations in sediment discharge, coupled with a right-skewed and heavy-tailed statistical distribution and the identification of 12.88% outliers, collectively suppressed the Nash–Sutcliffe efficiency (NSE) at this station. From a process-based perspective, the 3,700-fold variability in the sediment-to-flow discharge ratio corroborates the occurrence of hysteresis, underscoring the dependence of sediment yield on antecedent catchment states, particularly soil moisture conditions. Against this backdrop, Hybrid Model 1 (LMLWDC –CANFIS) achieved superior performance, attaining an NSE of 0.63 and markedly outperforming standalone empirical sediment rating curve (SRC) and artificial intelligence architectures. The robustness of this configuration stems from the strategic integration of empirically derived sediment rating-curve knowledge with the adaptive learning capacity of the Coactive Neuro-Fuzzy Inference System (CANFIS). By explicitly accommodating uncertainty and nonlinearity inherent in flow-sediment dynamics, the model establishes a coherent balance between numerical prediction and hydromorphological realism, thereby enhancing predictive reliability under highly variable sediment transport regimes.
Conclusion
The results unequivocally demonstrate that the synergistic integration of empirical hydrological reasoning and artificial intelligence provides a resilient and efficient modeling framework for watersheds characterized by pronounced skewness and a high prevalence of anomalous observations. In this context, the CANFIS architecture, leveraging fuzzy logic to resolve complex nonlinear interactions, substantially improves the representation of hysteretic sediment dynamics and offers clear advantages over conventional computational structures, including multilayer perceptron (MLP), general feed-forward (GFF), radial basis function (RBF), support vector machine (SVM), and self-organizing feature map (SOFM) networks. Given the strategic significance of flood events in water resources engineering, the rigorous interrogation of outlier behavior is not optional but imperative. The proposed hybrid framework effectively captures and interprets these statistical anomalies with acceptable precision, thereby reinforcing its applicability in suspended sediment transport management under extreme hydrological conditions. To advance methodological robustness and scientific depth in future investigations, the establishment of comprehensive, high-resolution databases—particularly those documenting peak flood discharges—is strongly recommended. Incorporating auxiliary predictors such as antecedent soil moisture, rainfall intensity, and vegetation cover indices would further strengthen process representation and address hysteresis phenomena. Moreover, deploying advanced machine-learning paradigms, including Long Short-Term Memory (LSTM) networks, alongside robust preprocessing techniques such as quantile regression and outlier treatment methods, has strong potential to attenuate systematic error and enhance model confidence in integrated watershed management applications. Additionally, ensemble methods combining multiple model outputs and uncertainty quantification frameworks are recommended to improve predictive reliability and decision-making support in sediment yield estimation.
Keywords
)