Machine learning applied to global scale species distribution models

Species Distribution Models (SDMs) are widely used in ecology to analyze historical and future patterns of marine species distributions. Given the growing impact of climate change, predicting potential shifts in species ranges has become a key challenge. In this study, we apply Bayesian Additive Regression Trees (BART), a non-parametric machine learning algorithm, to estimate and forecast the global distribution of marine turtle species under different climate change scenarios. We model both individual species and their combined functional group, assess their historical and future habitat suitability, and examine the contribution of key environmental predictors. To evaluate BART's performance, we conduct a simulation study under two contrasting distributional scenarios: a cosmopolitan and a persistent species. We also test the sensitivity of BART to pseudo-absence data and compare its performance with MaxEnt and Generalized Additive Models (GAMs). Results indicate that BART performs slightly better overall, particularly under pseudo-absence settings, showing higher accuracy and more stable sensitivity and specificity. These findings highlight BART as a reliable alternative for long-term, global-scale species distribution modeling in marine systems.