Machine Learning-Based Assessment of Soil Dispersibility and Erosion Types in Earthfill Dam: A Case Study from South Central Vietnam

The stability of earthfill dam foundations is crucial for ensuring a reliable water supply and protecting downstream infrastructure. Currently, assessments of soil dispersibility and erosion in earthfill dams primarily rely on empirical formulas and expert judgment. However, traditional formulas often fail to cover the full range of real-world classification scenarios and require the continuous development of new thresholds for each specific case, making management challenging. We propose using machine learning combined with field survey data to offer a new approach for classifying soil dispersity and erosion types. This study focuses on comparing machine learning-based classifications with traditional expert-based methods to predict soil dispersity and erosion types across dams in South-Central Vietnam. Machine learning models, particularly Random Forests, outperformed traditional methods in both accuracy (R² = 0.92 for soil dispersity and R² = 0.92 for erosion type) and consistency, especially in complex cases. The classifications generated by the machine learning model were consistent with traditional models while better addressing complex scenarios. Key predictors such as quartz content, moisture, and mineralogy were identified as important factors influencing soil behavior.