Ensemble learning and radial basis functions improve transferability of functional response models in species distribution modeling

Predictions from species distribution models (SDMs) often fail when transferred to new geographic regions or time periods, limiting their utility for biodiversity forecasting under environmental change. This lack of transferability stems from functional responses in habitat selection, where animals respond to the same habitat differently depending on the availability of alternative habitats. The Generalized Functional Response (GFR) framework explicitly models selection coefficients as functions of habitat availability, but existing polynomial implementations face a fundamental tradeoff: low-order polynomials are too rigid to capture complex responses, while high-order polynomials overfit and transfer poorly. We developed flexible extensions of the GFR framework that replace global polynomial functions with local radial basis functions (RBF-GFR) and combined both approaches with modern machine learning methods: classification and regression trees (CART), random forests (RF) and extreme gradient boosting (XGBoost). We systematically compared the out-of-sample predictive performance of 12 modeling approaches using block cross-validation across four contrasting datasets—two individual-based simulations with known ecological mechanisms, wolf telemetry data and sparrow colony surveys. Ensemble methods combining functional response frameworks with RF or XGBoost consistently ranked in the top three performers across all datasets. Out-of-sample R² scores improved substantially over traditional GLMs, with increases from 0.25 to 0.85 in individual cases and typical gains of 0.20–0.50. The RBF-GFR-RF model showed the most consistent transferability. Critically, ensemble averaging provided similar protection against overfitting as explicit regularization while achieving superior out-of-sample accuracy. Spatial predictions revealed that standard GLMs systematically under-predicted abundance hotspots, while unregularized flexible GFR models exaggerated extremes. Combining functional response theory with ensemble learning, particularly random forests, offers a practical path toward robust, transferable SDMs. Our comparative framework demonstrates that local basis functions paired with ensemble methods consistently outperform traditional approaches across diverse ecological systems and data types. The methods are computationally feasible for realworld applications and provide substantial improvements in predicting species distributions under novel environmental conditions, addressing a critical limitation in current SDM practice.