Format: ORAL

Csar Leblanc1,2, Pierre Bonnet2,Maximilien Servajean3,andAlexis Joly1

1 Inria, Zenith, Montpellier, France 2 CIRAD, AMAP, Montpellier, France 2 LIRMM, ADVANSE, Montpellier, France

Biodiversity is under severe pressure, as many different disturbance events threaten organisms with extinction and ecosystems with collapse. Therefore, species and habitat distribution modelling are increasingly relevant. They aim to quantify the statistical links between environmental covariates and respectively species or habitat occurrences. Herein, we introduce GeoPl@ntBERT, a novel framework to model species assemblages and produce distribution maps for individual habitats. It leverages a sophisticated deep learning pipeline based on computer vision (convolutional neural networks) and natural language processing (transformers). In particular, it focuses on (i) image classification (plants assemblages are created based on satellite images and rasterized environmental data), (ii) fill-mask (predicted plant species are then translated in context into a modelled ecological process) and (iii) text classification (habitat types are finally assigned to sentences describing species compositions). To train and evaluate the models, millions of heterogeneous presence-only records recently collected by citizens coupled with hundreds of thousands standardized presence-absence surveys from a large database of vegetation plots were used. The resulting dataset covers over ten thousand different European vascular plant species. All steps are validated using a spatial block hold-out procedure (to account for spatial autocorrelation) and reach a high accuracy, as assessed by expert judgement. GeoPl@ntBERT represents a cutting-edge approach, offering a powerful tool for advancing our understanding of biodiversity dynamics and supporting conservation efforts.