Landslide detection and susceptibility analysis: A case study in Pieng stream catchment, Son La province

Abstract

Landslides are a major natural hazard causing significant property damage and loss of life worldwide. In this study, an enhanced landslide inventory was developed for the Pieng Stream catchment (Son La Province) using Object-Based Image Analysis (OBIA) combined with field surveys. Twelve conditioning factors were used to model landslide susceptibility through four machine learning algorithms: extreme gradient boosting (XGB), random forest (RF), multi-layer perceptron (MLP) and logistic regression (LR). Additionally, model interpretation was supported by SHAP, MDA, and PDP analyses. The results demonstrated a high level of reliability for the OBIA method (TPR = 0.886, TS = 0.602). Among the tested models, the XGB model showed the best performance, achieving an AUC of 0.961, an F1 score of 0.915, and an accuracy of 0.915 on the testing dataset. The two most influential predictors identified were lineament density and aspect. An increase in landslide probability was observed with increasing slope, relative relief, lineament density, river density and aspect (0-150°). A total of 88% of testing landslide points were correctly classified within high to very high susceptibility areas, while areas outside the AOA covered merely 0.79% of the study region, indicating a high level of model applicability.