Current and future habitat suitability and bioactive potential of the carob tree (Ceratonia siliqua L.) across the western Mediterranean: a spatially validated ensemble and explainable machine-learning framework

The carob tree ( Ceratonia siliqua L.) is a drought-adapted evergreen legume of the Mediterranean Basin whose pods and leaves contain polyphenols of nutritional interest. Projections of its future climatic suitability remain scarce, and whether suitable areas coincide with areas of intense climatic stress is unknown. From 57,855 Global Biodiversity Information Facility records, 3957 presences thinned at 5 km were retained across Morocco, Spain, Portugal and Italy; of these, 3956 with complete bioclimatic data were modelled. Four algorithms (generalized linear models, generalized additive models, random forests, and extreme gradient boosting) were combined into a true-skill-statistic (TSS)-weighted ensemble, validated by 150-km spatial block cross-validation and projected with three climate models under SSP245 and SSP585 for 2050 and 2070. An exploratory Bioactive Potential Index (BPI) combined five climatic proxies of thermal and hydric stress using literature-based weights. The spatially cross-validated ensemble achieved AUC 0.893 ± 0.016, TSS 0.635 ± 0.032 and continuous Boyce index 0.975 ± 0.017; random cross-validation of the random forest inflated AUC to about 0.985. Mean annual temperature was the leading predictor across three importance metrics. Suitable area increased under SSP245 (+ 17.6% by 2050; +21.5% by 2070) and was nearly unchanged under SSP585 by 2070 (− 0.02%), despite extensive spatial turnover and a 190-km poleward shift of the distribution centroid. Environmental novelty rose from 57.2% to 64.7% with scenario severity. BPI values were highest in Morocco, increased with warming everywhere and were stable under ±20% weight perturbation (Spearman ρ ≥ 0.989); the country ranking held within suitable cells, but the between-country range narrowed from 12.3 to 5.0 units. At presence sites, no association between suitability and BPI was detected after correction for spatial autocorrelation (Pearson r = −0.07, p = 0.60, effective df ≈ 51). Coastal and insular areas remained suitable under all scenarios and are candidate climate refugia. Since the BPI shares its climate inputs with the distribution model and was not calibrated against phytochemical data, it indicates where climatic stress is more intense, but not where polyphenol content is higher; field sampling along the aridity gradient is needed to test it.