Interpretable metaheuristic-optimized machine learning for predicting biodiversity indices in the Hyrcanian forest

Operational biodiversity indicators are essential for tracking ecological change and evaluating forest management effectiveness. However, predicting multiple biodiversity indices remains challenging because their associations with environmental and stand-structural conditions are nonlinear, context-dependent, and index-specific. We developed an interpretable machine learning framework that integrates metaheuristic optimization, multi-index biodiversity assessment, and model-based ecological interpretation using data from 811 national forest inventory plots across the Hyrcanian forest of northern Iran. Species richness, Shannon–Wiener diversity, Gini–Simpson diversity, inverse Simpson diversity, and Pielou's evenness were derived from tree-level species composition, with basal-area proportions used where appropriate for the abundance-based indices. Topographic, stand structural, edaphic, and bioclimatic variables were used as predictors. Random forest (RF), support vector machine (SVM), and k-nearest neighbors (kNN) were independently optimized for each biodiversity index using four population-based metaheuristic algorithms to select model-specific hyperparameters by minimizing validation RMSE; final model performance was evaluated on held-out test data using R 2 , RMSE, and MAE. Predictive performance differed substantially among biodiversity indices: optimized SVM performed best for Shannon–Wiener, Gini–Simpson, and inverse Simpson diversity, with test R 2 values of 0.7408, 0.8472, and 0.7387, respectively, whereas RF achieved the highest test R 2 for species richness and Pielou's evenness, with values of 0.4125 and 0.2670, respectively. Thus, abundance-weighted diversity indices were considerably more predictable than richness and evenness, highlighting the need for index-specific predictive models rather than a single universal modeling approach. SHapley Additive exPlanations (SHAP), partial dependence, and interaction analyses consistently identified elevation as the dominant predictor across biodiversity indices, whereas diameter at breast height, basal area, and slope showed distinct nonlinear response patterns. By combining index-specific hyperparameter optimization with ecological interpretation, the proposed framework supports the comparison and interpretation of biodiversity predictions for monitoring and ecological indicator development in structurally complex forests.