Assessment of data-driven models for estimating total sediment discharge

Abstract

Estimating total sediment discharge is challenging. This study aims to assess performances of various data-driven models including empirical equations, machine learning (ML), and ensemble models for such estimations. The ML models include Support Vector Machine (SVM), Artificial Neural Network (ANN), K Nearest Neighbor (KNN), Random Forest (RF), and Decision Tree Regression (DTR). For this purpose, 543 widely-ranged data were collected from the United States Geological Survey (USGS) resources and used to train and test different models. Ranking different models demonstrated that Ackers and White's equation outperformed multiple linear regression (MLR) and SVM, which indicates that all ML models do not necessarily outperform empirical equations. Moreover, despite conducting multiple runs and parameter tuning, the results consistently indicated that increasing the number of hidden layers and neurons in ANN structures did not significantly improve the overall performance of the ANN models. In addition, the nonlinear ensemble model outperformed all methods and placed first in the ranking. Despite a notable difference between metrics obtained by KNN for the train and test data, it outperformed other methods and ranked second, while ANN achieved the third-best ranking place. The obtained result was also confirmed by the reliability analysis and confidence limits. However, due to negative predictions for some small sediment discharges by the nonlinear ensemble method, it did not demonstrate good reliability. Finally, the comparative analysis indicates that selecting a suitable model for estimating sediment discharges with a desirable accuracy is challenging, while further studies are required to assess other ML models or variants of ensemble models.