Machine Learning for post-mining ground stability: predicting sinkhole geometry to support risk management and land reuse

Post underground mining activities can trigger sinkholes that pose long-term risks to infrastructure, public safety, and the sustainable redevelopment of former coal and mining regions. Robust, scalable tools for estimating sinkhole geometry are therefore essential for post-mining risk management, land-use planning, and evidence-based remediation strategies. This study investigates the use of supervised machine learning methods to predict key sinkhole size parameters associated with abandoned underground workings. A multi-country European database was compiled, integrating sinkhole inventories with geological and mining attributes, including characteristics of underground excavations and overburden conditions. Several regression algorithms were trained and compared to estimate sinkhole geometry from available predictors, and model performance was evaluated using standard statistical metrics. To support interpretability and practical adoption, feature-importance analyses were performed to identify the most influential factors controlling predicted sinkhole dimensions. The results demonstrate the potential of data-driven modelling to enhance post-mining hazard assessment and to inform prioritization of monitoring, remediation, and safe land reuse, contributing to risk-aware, sustainable transition pathways in regions affected by underground mining.