Machine learning-based framework for forecasting mass-mortality events in Mediterranean coralligenous habitats
Maria Del Mar Bosch-Belmar
1,2,3
Francesco Paolo Mancuso
1,2
Mario Francesco Tantillo
1
1
Deparment of Earth and Marine Science (DiSTeM), University of Palermo, Via Archirafi, 22, Palermo, PA - 90123, Italy
2
, National Biodiversity Future Center, Piazza Marina, 61, Palermo, PA - 90133, Italy
3
, Consorzio Nazionale Interuniversitario per le Scienze del Mare (CoNISMa), Piazzale Flaminio, 9, Rome, RO - 00196, Italy
Global warming is triggering increasingly frequent mass-mortality events (MMEs) in Mediterranean benthic habitats. Documented MMEs involve coralligenous communities, reef-building assemblages that have recurrently collapsed since the 1980s. Occupying a narrow thermal range, these habitat-formers are highly susceptible to extreme marine heatwaves and environmental anomalies. Although inclusion within Marine Protected Areas (MPAs) reduces local pressures, their capacity to buffer climate-driven mortality is uncertain. Moreover, the environmental combinations that amplify or dampen coralligenous MME severity remain largely qualitative. Our primary objective is to elucidate the environmental drivers that either amplify or mitigate the transition towards severe mass-mortality events in Mediterranean coralligenous communities, and to assess the potential of protective measures in managing these induced changes. To this end we apply explainable artificial-intelligence methods, specifically eXtreme Gradient Boosting (XGBoost) models, and assess their classification skill to judge whether the resulting models can underpin an early-warning tool for the management and conservation of these habitat-forming organisms. Using the Mediterranean MME database provided by Carlot et al. (2025), we combined monthly oceanographic variables from Copernicus Marine Environment Monitoring Service and annual marine-heat-wave metrics with site descriptors such as depth, geolocation and MPA protection. We fit multiple XGBoost models using train-test partitions that combine complementary spatial-temporal validation strategies. Model behaviour is interpreted with SHapley Additive exPlanations (SHAP), which decompose each prediction into additive variable contributions. Performance metrics are employed to quantify explanatory power. Preliminary results indicate that the best XGBoost models, trained with cross-validated hyperparameter optimization, achieved robust discrimination of severe MME years (AUPRC: 0.78, F1: 0.69, LOYO validation). Predictive accuracy was higher under random KFold splits (AUPRC: 0.85), but declined when forecasting years not seen during training, underscoring the challenge of extrapolating to novel environmental regimes. These findings suggest that early-warning predictions are feasible and can identify a substantial proportion of severe events, although further validation on fully independent years and real-time scenarios remains necessary.
Effetti del disturbo sui sistemi ecologici