From Prediction to Decision: An Explainable AI Framework for Urban Ground Subsidence Risk Management

Urban ground subsidence is a rare but hazardous disaster driven by aging infrastructure, dense underground utilities, and groundwater dynamics. Its prevention is challenging due to infrequent occurrence and complex, nonlinear interactions. While machine learning (ML) has been widely applied to subsidence risk prediction, most studies focus on predictive accuracy or risk visualization, with limited efforts to translate results into practical maintenance and investigation decisions. This study proposes an explainable AI-based framework that converts subsidence risk predictions into actionable management strategies. A dataset integrating subsidence history and underground utility attributes was constructed using 500 m × 500 m grid units. An XGBoost classification model with a One-vs-Rest structure was applied for three-class risk management. Considering class imbalance, the model achieved a macro-F1 score of 0.62, a ROC–AUC of 0.81, and a recall of 0.74 for the high-risk class. To enhance interpretability, SHapley Additive exPlanations (SHAP) were employed to quantify feature contributions by risk class. Underground utility density showed the strongest overall influence on predictions. Pipe age exhibited nonlinear effects and interactions with density. Reorganizing age into Young (≤10 years), Mid (10–20 years), and Old (≥20 years) groups improved interpretability, with the Old group contributing most strongly to high-risk predictions. Older pipelines were associated with higher predicted probabilities of high-risk, whereas younger pipelines showed lower probabilities. By separating prediction, interpretation, and decision-making stages, the proposed framework provides clear criteria for inspection prioritization and maintenance planning, supporting transparent and evidence-based urban subsidence risk management. • Explainable AI links subsidence prediction to management decisions • The XGBoost model classifies three urban subsidence risk levels • SHAP identifies density and pipe age as key risk drivers • The framework turns AI outputs into inspection and maintenance priorities • It is designed for rare and highly imbalanced disaster data