Abstract Sediment settling velocity (SSV) is one of the most important parameters in sediment transport studies. Accurate estimation of SSV is, thus, of great significance for river basin planning and management. Many factors influence SSV in highly complex and nonlinear ways, which make accurate estimation of SSV a very challenging task. To better estimate SSV, in the present study, three machine learning models, namely Feed Forward Neural Network (FFNN), Deep Learning (DL), and Decision Tree (DT), are developed. Data from the previous literature for sand and gravel classes, including nominal diameter of the sediment, kinematic viscosity of the fluid, submerged specific gravity of the sediment, and observed SSV are used as inputs to these models. To assess the superiority of these models against traditional methods, if any, the modeling results are compared with four common SSV estimation formulas and also with the results from a genetic programming model previously developed. The results show that the DT model outperforms all the conventional formulas and the genetic programming model, as well as the FFNN and DL models, for both the sand and gravel classes. The DL method does not perform well when the data size is small, and the results are even worse than those from the FFNN model. The results from this study are certainly encouraging regarding the suitability and effectiveness of machine learning models for reliable estimation of SSV, provided the limitations about the data are properly understood.

中文翻译：

---

估算沉积物沉降速度的机器学习方法

摘要 沉积物沉降速度（SSV）是沉积物迁移研究中最重要的参数之一。因此，准确估算SSV对于流域规划和管理具有重要意义。许多因素以高度复杂和非线性的方式影响 SSV，这使得准确估计 SSV 成为一项非常具有挑战性的任务。为了更好地估计 SSV，在本研究中，开发了三种机器学习模型，即前馈神经网络 (FFNN)、深度学习 (DL) 和决策树 (DT)。先前文献中关于沙子和砾石类别的数据，包括沉积物的公称直径、流体的运动粘度、沉积物的淹没比重和观测到的 SSV，被用作这些模型的输入。为了评估这些模型相对于传统方法的优越性，如果有的话，建模结果与四种常见的 SSV 估计公式以及先前开发的遗传编程模型的结果进行了比较。结果表明，无论是沙子还是砾石类，DT 模型都优于所有常规公式和遗传规划模型，以及 FFNN 和 DL 模型。DL 方法在数据量较小时表现不佳，结果甚至比 FFNN 模型更差。如果正确理解数据的局限性，这项研究的结果肯定会令人鼓舞，机器学习模型在可靠估计 SSV 方面的适用性和有效性。结果表明，无论是沙子还是砾石类，DT 模型都优于所有常规公式和遗传规划模型，以及 FFNN 和 DL 模型。DL 方法在数据量较小时表现不佳，结果甚至比 FFNN 模型更差。如果正确理解数据的局限性，这项研究的结果肯定会令人鼓舞，机器学习模型在可靠估计 SSV 方面的适用性和有效性。结果表明，无论是沙子还是砾石类，DT 模型都优于所有常规公式和遗传规划模型，以及 FFNN 和 DL 模型。DL 方法在数据量较小时表现不佳，结果甚至比 FFNN 模型更差。如果正确理解数据的局限性，这项研究的结果肯定会令人鼓舞，机器学习模型在可靠估计 SSV 方面的适用性和有效性。