불균형 다변량 데이터를 활용한 인공지능 기반 지반함몰 예측 성능 분석 및 변수 중요도 해석

초록

This study evaluates the probability of ground subsidence and examines the applicability of AI-based prediction models using publicly available data. Gwangju Metropolitan City was selected as the study area, and a spatial dataset was constructed using five years of ground subsidence records (2019–2023), rainfall data (including 7-day and 30-day cumulative rainfall), geological information, and the distribution of aged buildings. To address severe class imbalance, techniques including undersampling, the synthetic minority oversampling technique, class weighting, and focal loss were applied. A comparative analysis was then performed using Random Forest, XGBoost, and deep neural networks. Although overall F1-scores were constrained by the intrinsic rarity of subsidence events, precision–recall analysis demonstrated that the models achieved meaningful classification performance. Among the evaluated models, XGBoost exhibited the most stable and accurate predictive capability. Feature importance analysis identified cumulative rainfall and the number of aged buildings as the most influential variables, indicating that urban ground subsidence is predominantly governed by the combined effects of long-term rainfall accumulation and aging underground infrastructure.

키워드