Machine Learning in Biodiversity Forecasting

Abstract

Machine learning has transformed the quantitative toolkit of biodiversity science, enabling predictions of species distributions, population trends, and ecosystem change at spatial and temporal scales that were previously infeasible with traditional statistical methods. The application of deep learning to remote sensing data enables habitat mapping at sub-meter resolution globally; recurrent neural networks and transformers applied to long-term monitoring data enable population trend forecasting beyond the end of the monitoring record; and graph neural networks incorporating species interaction networks enable prediction of extinction cascades from primary species losses. Yet the proliferation of ML methods in biodiversity science has outpaced systematic benchmarking of their performance relative to classical statistical approaches and against each other -- creating a field where the 'newest' method is often adopted before it is confirmed as better than established alternatives. This review and benchmark study systematically evaluates 18 ML methods across six biodiversity forecasting tasks using a standardised holdout evaluation framework with 247 independent test datasets. Deep learning methods outperformed classical statistical methods by a mean 18.4% in prediction accuracy across all tasks -- but only when training data exceeded critical thresholds (> 1,000 records for SDM; > 20 years for trend forecasting). A practical decision framework for ML method selection across biodiversity forecasting contexts is developed, identifying when ML adds genuine predictive value over simpler approaches.