A Comparative Evaluation of Deep Learning and Machine Learning Models for River Suspended Sediment Concentration Forecasting

Abstract

Suspended sediment concentration (SSC) in rivers is a crucial parameter required in hydrological studies, water resources management, and many other relevant applications. This study presents a comparative assessment of deep learning (DL) and machine learning (ML) methods in river SSC prediction of two river stations on the Mississippi River, United States. To that end, two single DL models, namely recurrent neural networks (RNN) and bidirectional RNN (BiRNN) were developed. Generally, the RNN was found to outperform the BiRNN for predicting SSC. Furthermore, a convolutional neural network (CNN) was coupled on the applied DL models to create the hybrid RNN-CNN and BiRNN-CNN models. The results denoted that the BiRNN-CNN models generally performed better compared with RNN-CNN ones. Besides the four types of DL models, three forms of ML models, including adaptive boosting (AdaBoost), natural gradient boosting (NGBoost), and gradient boosting regression trees (GBRT) were also established. As a general conclusion, NGBoost and GBRT demonstrated the highest and lowest level of accuracy in river SSC forecasting. Eventually, the influence of input predictors on the outputs of models was done considering local interpretable model-agnostic explanations (LIME). Assessing the LIME outcomes for the selected samples of the test data revealed that the current daily river streamflow and one daily lagged SSC data were the most effective inputs on SSC prediction results.