Risk assessment of tunnelling-induced hydrogeological interference on springs using a machine learning approach

Tunnel excavation through mountain ranges can alter groundwater regimes, affecting spring discharge and dependent ecosystems. Anticipating these interferences is essential for environmental risk management, cost–benefit analysis, and effective mitigation planning. Conventional assessment methods are largely parametric, not physically based, and heavily dependent on expert judgment, introducing subjectivity and uncertainty. This study proposes a data-driven approach based on machine learning (ML), specifically a Random Forest (RF) model, to evaluate hydrogeological risk and spring vulnerability during the preliminary design phase of tunnels. The method exploits geological information typically available in early stages and is validated using two well-documented Italian case studies: The Bologna–Florence high-speed railway line, excavated in fractured sedimentary formations of the Northern Apennines, and the Gran Sasso highway tunnels, driven in karstified carbonate rocks of the Central Apennines. Both datasets are published, hydrogeologically validated, and supported by detailed pre- and post-excavation monitoring, ensuring robustness and consistency. The ML framework demonstrates strong predictive performance across standard classification metrics, including accuracy, precision, recall, F1-score, Matthews correlation coefficient (MCC), and area under the curve (AUC). While it employs input parameters comparable to those of parametric methods such as the drawdown hazard index (DHI), it differs by deriving parameter weights directly from data. Moreover, the use of Shapley additive explanations (SHAP) values enhances interpretability, mitigates black-box behaviour, and allows expert knowledge to be effectively integrated. The innovation, therefore, lies in offering a transferable and reproducible tool to support early-stage tunnelling decisions, representing a clear improvement over traditional qualitative or semi-quantitative approaches.