Predictive Approaches for Ground Subsidence Hazard Using Machine Learning

This study developed a machine learning-based predictive model using attribute data from underground pipelines and historical ground subsidence records to forecast susceptibility in urban areas. The study area was discretized into a 100m×100m grid, and datasets were created by integrating pipeline attributes and corresponding ground subsidence data. Key input parameters, such as pipeline diameter, usage duration, and density, were chosen based on pipeline length, while the ground subsidence risk level was established as the output variable using historical data. The datasets were divided into training and evaluation sets. Evaluation metrics included F1-score and AUC. Comparative analysis showed that the RandomForest classifier achieved the best performance (F1-Score: 0.430, AUC: 0.820) when applied to a preprocessed dataset with a 5-year pipeline utilization, 100mm diameter, and 1–2 ground sinks in the Level 1 risk range, demonstrating the mapping approach's efficacy in predicting ground subsidence risk in urban areas.