GeoPl@ntNet: A Deep Learning Workflow for Mapping European Plant Species and Habitats at Very-High Resolution
Date:
Link to the meeting: click here
We introduce GeoPl@ntNet, a deep-learning based workflow aiming at mapping European plant species (over 10,000 organisms) and ecosystems (over 200 EUNIS habitats) at very-high resolution (50m) and deriving biodiversity indicators (e.g., species richness and diversity, presence of protected or threatened species, and number of invasive species). The pipeline is based on computer vision (convolutional neural networks) and natural language processing (transformers) and uses millions of heterogeneous presence-only records coupled with hundreds of thousands standardized presence-absence surveys. In particular, it focuses on (i) image classification (plant assemblages are created with satellite images and rasterized environmental data), (ii) fill-mask (predicted species are translated into a modelled ecological process) and (iii) text classification (habitats are assigned to sentences describing species compositions). We will discuss the validation and interpretability of the results as well as the potential benefits and risks of GeoPl@ntNet as a powerful tool for understanding and monitoring biodiversity dynamics across Europe.