Optimizing Machine Learning Models for Landslide Susceptibility Mapping in Yen Bai Province, Vietnam

Abstract

Landslide hazard assessment is crucial for managing and mitigating landslide risks. Landslide Susceptibility Mapping (LSM) provides a practical and cost-effective tool for zoning areas prone to landslides. LSM expresses in the form of a probability of landslide risk in each pixel.  This study implements Machine Learning (ML) models to generate Susceptibility Maps in Van Yen (VY), making them applicable to diverse topographic regions, particularly in areas significantly affected by human activities. Landslides were mapped after landslide events occurred. The datasets were created by combining balanced numbers of landslide and non-landslide points with 17 contributing factors, including topographic, geological, hydrological, anthropogenic, and vegetation factors. To ensure all factors have the same range of values, landslide areas standardized using the Frequency Ratio (FR). The effectiveness of the method will be evaluated on the same dataset before and after applying FR. Four ML models Random Forest (RF), Support Vector Machine (SVM), Logistic Regression (LR), and Extreme Gradient Boosting (XGBoost) were trained in VY area, then validated on the dataset in Mu Cang Chai (MCC). Evaluation process was conducted using Accuracy Evaluation and Efficient Global Optimization (EGO). The results indicate that RF and XGBoost achieved the highest and most consistent performance, with the highest learning capacity. However, according to the statistical evaluation, Geologic factors have low influence on classification, suggesting the need for including other critical geologic factors. The study can act as future framework for susceptibility analysis on python environment.