An Optimization-Focused Comparative Framework for Machine Learning-Based Suspended Sediment Load Modeling

Abstract Urbanization in the Göksu River Basin increases sediment movement toward the Göksu Delta. The research used hydrometeorological data to assess machine learning techniques, including hybrid and deep learning systems, for Suspended Sediment Load (SSL) forecasting. The development of conventional models, including Backpropagation Neural Network (BPNN), Radial Basis Function Neural Network (RBFNN), and Support Vector Machine (SVM), received optimization algorithm improvements through the implementation of Adaptive Moment Estimation (ADAM), Sequential Minimal Optimization (SMO), and the Water Cycle Algorithm (WCA), which boosted model accuracy and development speed. The BPNN_ADAM hybrid model performed better than all other methods because it achieved superior R 2 , adjusted R 2 , Nash–Sutcliffe Efficiency (NSE), RMSE-observations standard deviation ratio (RSR), and Normalized Root Mean Square Error (NRMSE) results, resulting in a low Mean Absolute Error (MAE) of 0.09. The BPNN_WCA hybrid system showed consistent performance across all tested metrics, which measured both error and operational efficiency. The deep learning models Long Short-Term Memory (LSTM) and Bidirectional LSTM (Bi-LSTM) achieved exceptional predictive performance, with R 2 values close to 0.99 in both univariate and multivariate testing modes. Deep learning models (LSTM/Bi-LSTM) have NRMSE values of 8–12%, followed by optimized hybrid models such as BPNN_ADAM and BPNN_WCA, with NRMSE values of 35–50%. In contrast, standalone machine learning models, such as SVM, show significantly higher errors with NRMSE values exceeding 60–90%. The research indicated that time-lagged hydrological and sediment variables play a crucial role in improving SSL prediction accuracy while also capturing the memory of catchment systems.