Quantitative Identification of Governing Factors for Sediment Transport Using Machine Learning

다국어 초록 (Multilingual Abstract)

Sediment transport plays a critical role in river morphology and directly influences the stability of hydraulic structures. Excessive or uncontrolled sediment movement can lead to engineering issues such as pier scouring, reservoir sedimentation, and structural failure. Despite the importance of sediment transport prediction, conventional physics-based models often derived from controlled laboratory experiments struggle to generalize to complex natural conditions and provide limited insights into variable interactions. Additionally, there remains no consensus on the dominant variables affecting sediment concentration. To overcome these limitations, this study applies machine learning techniques to identify dominant variables influencing sediment concentration. The central hypothesis is that variables consistently associated with high predictive accuracy under identical modeling conditions can be considered dominant. A combination of statistical, structural, and physical perspectives was employed to achieve this goal.
Four predictive models—Random Forest, Artificial Neural Network, Support Vector Machine, and Linear Regression—were constructed using identical input conditions. Input variables were categorized into three types: dimensional variables obtained from measurements, dimensionless variables derived by dimensional analysis, and principal components from principal component analysis. Model performance was assessed using RMSE, R², Correlation coefficient, and the within ratio (predicted-to-observed ratio between 0.5 and 2).
Among the models, Random Forest demonstrated the best and most stable performance across different sediment concentration levels and was thus selected for further analysis. Among input types, the model using dimensionless variables outperformed the others, highlighting their strength in removing scale effects and improving generalizability.
To identify dominant variables, three evaluation methods were used: composite score analysis across all variable combinations, SHAP (Shapley Additive Explanations) values, and Random Forest derived feature importance. Dimensionless unit stream power consistently ranked highest across all methods. Its physical characteristics integrate flow energy and particle settling dynamics, reinforcing its role as a key variable.
This study also emphasizes that statistical or structural dominance (e.g., high correlation or PCA loading) does not guarantee physical relevance. Thus, physical interpretability was used as a critical criterion in variable selection, ensuring that identified dominant variables align with real-world sediment transport mechanisms.
The findings contribute to improving the scientific basis of sediment prediction by proposing a systematic, data-driven approach to identifying physically dominant variables. While this study was based on laboratory data, future work will focus on validating the models using field data to enhance their reliability and applicability to river systems.

목차 (Table of Contents)

• Abstact i
• Table of Contents iii
• List of Tables vi
• List of Figures vii
• Chapter 1. Introduction 1
• 1.1. Background 1
• 1.2. Objectives 3
• 1.3. Overview 4
• Chapter 2. Literature review 7
• 2.1. Conventional approach for sediment transport 7
• 2.1.1. Quantitative approaches to sediment transport mechanisms . 11
• 2.1.2. Application of sediment transport prediction 13
• 2.2. Application of machine learning 15
• 2.2.1. Application of machine learning to hydraulic engineering 15
• 2.2.2. Application of machine learning to bed morphology 16
• 2.2.3. Application of machine learning to sediment transport 17
• 2.2.4. Determination of dominant factors using machine learning .. 22
• Chapter 3. Theoretical background 23
• 3.1. Sediment transport analysis 23
• 3.1.2. Shear stress-based method 25
• 3.1.3. Energy-based method 26
• 3.1.4. Discharge and velocity-based method 28
• 3.1.5. Probabilistic method and Others 29
• 3.2. Machine learning 30
• 3.2.1. Artificial Neural Network (ANN) 31
• 3.2.2. Random Forest (RF) 34
• 3.2.3. Support Vector Machine (SVM) 35
• 3.3. Hyperparameter optimization 36
• 3.3.1. Grid Search 36
• 3.3.2. Random Search 37
• 3.3.3. Bayesian Search 38
• 3.3.4. Selection of hyperparameter optimization strategy 39
• 3.4. Linear regression 41
• 3.5. Dimensional analysis 42
• 3.6. Principal Component Analysis (PCA) 44
• Chapter 4. Methodology 46
• 4.1. Dataset construction 46
• 4.1.1. Data collection 46
• 4.1.2. Data preprocessing 53
• 4.2. Experimental design 54
• 4.2.1. Derivation of dimensionless variables 54
• 4.2.2. Preliminary analysis 60
• 4.2.3. Principal Component Analysis (PCA) 62
• 4.2.4. Definition of model performance evaluation metrics 64
• 4.2.5. Determination of dominant factor 65
• 4.2.6. Hyperparameter optimization 67
• Chapter 5. Results and discussion 70
• 5.1. Selection of optimal model 71
• 5.2. Predictive performance with respect to the type of variables 73
• 5.3. Determination of dominant variable 76
• 5.3.1. Composite score 76
• 5.3.2. Shapley value. 78
• 5.3.3. Feature importance 80
• 5.3.4. Identification of governing variable 82
• Chapter 6. Conclusion 90
• Reference 93
• 국문초록 100