Building AI Training Data for Landslide Susceptibility Assessment Based on Geospatial Information

Abstract

Landslides cause significant damage to human lives, infrastructure, and property, with their frequency and intensity increasing due to recent climate change and land use changes. Landslide susceptibility assessments evaluates vulnerable areas based on environmental factors such as soil properties and vegetation that influence landslide occurrence, providing essential tools for risk mitigation. Artificial intelligence (AI) techniques, including machine learning and deep learning, are employed to identify relationships between various environmental factors and landslide occurrence patterns. Since AI-based landslide susceptibility analysis results are used for decision-making in landslide damage reduction, ensuring high accuracy is crucial. While data quality significantly affects the accuracy of landslide susceptibility analysis in AI model training, studies comprehensively addressing the complete data curation workflow from data collection to model input variable selection remain limited. Therefore, this study developed training datasets for landslide susceptibility analysis in Chungcheongbuk-do Province, Republic of Korea, where massive landslides occurred in 2020. The comprehensive methodology, including raw data collection, data preprocessing, and variable selection for acquiring data on soil, vegetation, land cover, geology, and topography that influence landslide occurrence, is applicable not only to the current study area but also to other landslide-susceptible areas. Consequently, this study provides data construction guidelines for AI-based landslide susceptibility analysis, contributing to improved accuracy in future landslide susceptibility assessments.