GeoPl@ntNet: A Remote Sensing-Based Deep Learning Workflow for Biodiversity Mapping and Monitoring
Date:
Link to the conference: click here
Biodiversity is under pressure due to a variety of environmental disturbances, making its monitoring essential for effective conservation action. Herein, we present GeoPl@ntNet, an advanced satellite remote sensing (SRS) and deep learning-based workflow designed to map and monitor European plant species (over 10,000 organisms) and ecosystems (over 200 EUNIS habitats) while providing biodiversity indicators, all at very-high resolution (50m). GeoPl@ntNet leverages both computer vision (convolutional neural networks) and natural language processing (transformers) to integrate multiple biodiversity and environmental data streams, using millions of heterogeneous presence-only records combined with hundreds of thousands of standardized presence-absence surveys. The framework is composed of three components: (i) image classification, where satellite imagery (i.e., patches and time series) and environmental rasters (e.g., bioclimatic rasters and soil rasters) are used to predict plant assemblages; (ii) fill-mask modeling, which gets a syntaxic understanding of vegetation patterns; and (iii) text classification, which uses the predicted assemblages to identify habitat types. These tasks enable GeoPl@ntNet to produce very high-resolution maps of individual species and habitats across Europe, and derive key biodiversity metrics, including species richness, presence of invasive or threatened species, and ecosystem health indicators. In addition, we will discuss the validation of all steps (i.e., the spatial block hold-out approach to address spatial autocorrelation), the interpretability of the maps (i.e., how they can offer insights into the dynamic interactions between environmental drivers and biodiversity patterns), and the results obtained (i.e., our model outperforming MaxEnt and expert systems). Finally, we will dive into the potential of GeoPl@ntNet as a powerful tool for understanding and monitoring biodiversity dynamics and see if the integration of SRS technologies and deep learning can enable us to enhance our comprehension of ecosystems. We will reflect on how it could help guiding conservation efforts and supporting policy frameworks aimed at reversing biodiversity loss in Europe.