Abstract

Abstract

Soil fungal communities are critical drivers of terrestrial ecosystem function, yet their global distribution remains largely unknown due to the challenges of widespread physical sampling. We developed a machine learning pipeline to predict fungal biodiversity across Europe and Asia using high-resolution, temporal satellite imagery. We introduce a novel feature set derived from a self-supervised learning (SSL) model applied to Sentinel time series. We trained a model on roughly 12,000 mycorrhizal fungal richness samples, comparing the predictive power of our SSL features against standard environmental datasets. Our combined model achieves a robust R2 of 0.53-0.55 across 50 cross-validation runs. We show that the SSL features are the single most important predictor group, outperforming traditional datasets and implicitly capturing land cover information. Furthermore, we demonstrate that prediction errors are geographically clustered in sparsely sampled regions, providing a data-driven method for identifying “biodiversity data deserts” and guiding future sampling efforts. This work presents a scalable framework for monitoring an overlooked component of global biodiversity and demonstrates the viability of temporally-rich, self-supervised representations for ecological modeling.

Bio

Robin Young is a first-year PhD student in Computer Science at the University of Cambridge.