Comparing Five Machine Learning Approaches for Local-Scale Landslide Susceptibility Mapping in the Huoi Reng Watershed, Central Vietnam

Abstract

Landslide susceptibility mapping supports disaster-risk reduction in steep tropical catchments where rainfall, deeply weathered materials, lithology and land-cover disturbance interact. This study compares artificial neural network (ANN), J48 decision tree, Random Forest (RF), Bagging with Random Forest base learner (Bagging-RF) and Support Vector Machine (SVM) models for the Huoi Reng watershed, Central Vietnam. The inventory contains 319 mapped landslide locations. Calibration used pre-2023 landslides, while 39 landslides mapped in 2024 and six field-checked landslides from 2025 were kept for a limited occurrence-only plausibility assessment. Presence samples were extracted from landslide cells on a common 30 m grid, and non-landslide samples were randomly selected outside mapped landslides and a three-cell exclusion buffer. Model comparison was revised from a cell-level random split to a five-fold spatial block validation design. Bagging-RF obtained the highest mean ROC-AUC (0.807 ± 0.063), followed closely by RF (0.805 ± 0.063); the difference between these two ensemble models was not statistically significant. Bagging-RF and RF also produced the lowest Brier scores (0.184 ± 0.027 and 0.187 ± 0.028), indicating smaller fold-level probability error than ANN, J48 and SVM. The 45 later landslide points were used only as an occurrence-only class-concentration check, not as a full validation set. RF captured 61.85% of the later landslide signal within 9.89% of the area mapped as high or very high susceptibility. The results indicate that tree-based ensembles provide the most stable local screening products, but the maps should be used for field-priority zoning rather than quantitative risk prediction until independent stable-site sampling and event-based rainfall validation are available.