Modeling habitat suitability for endangered herb (Salvia leriifolia Benth) using Innovative hybrid machine learning algorithms

Mapping habitat suitability is critical for conserving endangered medicinal and aromatic plants (MAPs) in degraded ecosystems. This study evaluated habitat suitability for Salvia leriifolia, an endangered herb in Khorasan Razavi, Iran, by comparing individual machine learning models (Random Forest [RF], Boosted Regression Tree [BRT], Generalized Linear Model [GLM]) with hybrid ensemble models (FDA-GLM-MDA and RF-CART-BRT). Using GIS, we compiled a dataset of 23 environmental and anthropogenic variables. Variable importance was assessed via Elastic Net (ENET), while model performance was evaluated using ROC-AUC metrics. Results identified distance from roads, calcium (Ca), organic matter (OM), and potassium (K) as the most influential predictors of habitat suitability. The individual RF model (AUC = 0.983) and hybrid RF-CART-BRT model (AUC = 0.970) demonstrated the highest predictive accuracy, underscoring the efficacy of tree-based and ensemble approaches in ecological modeling. However, while high-accuracy models offer precise predictions, their complexity may challenge practical application by conservation practitioners. This highlights the need to balance technical precision with operational feasibility in conservation planning. Although focused on Khorasan Razavi, the methodology is transferable to regions facing similar ecological pressures. By identifying key drivers of habitat suitability, this work enabled scalable strategies for protecting endangered MAPs. These insights offered actionable strategies for resource managers and agricultural planners to sustainably protect and cultivate S. leriifolia, while addressing predictive model implementation challenges.