Landslide susceptibility assessment of South Korea using stacking ensemble machine learning

Abstract

Landslide susceptibility assessment (LSA) is a crucial indicator of landslide hazards, and its accuracy is improving with the development of artificial intelligence (AI) technology. However, the AI algorithms are inconsistent across regions and strongly dependent on input variables. Additionally, LSA must include historical data, which often restricts the assessment to the local scale and single landslide events. In this study, we performed an LSA for the entirety of South Korea. A total of 30 input variables were constructed, consisting of 9 variables from past climate model data MK-PRISM, 12 topographical factors, and 9 environmental factors. Sixteen machine learning algorithms were used as basic classifiers, and a stacking ensemble was used on the four algorithms with the highest area under the curve (AUC). Additionally, a separate assessment model was established for areas with a risk of landslides affecting areas larger than 1 ha. The highest-performing classifier was CatBoost, with an AUC of ~ 0.89 for both assessments. Among the input variables, distance of road, daily maximum precipitation, digital elevation model, and soil depth were the most influential. In all landslide events, CatBoost, lightGBM, XGBoost, and Random Forest had the highest AUC in descending order; in large landslide events, the order was CatBoost, XGBoost, Extra Tree, and lightGBM. The stacking ensemble enabled the construction of two landslide susceptibility maps. Our findings provide a statistical method for constructing a high-resolution (30 m) landslide susceptibility map on a country scale using diverse natural factors, including past climate data.