L'intelligence artificielle au service de la modélisation de la biodiversité : de la prévision des assemblages d'espèces végétales à la compréhension et à la cartographie des habitats

Biodiversity is undergoing rapid change due to global environmental pressures, yet our ability to monitor and predict species distributions and ecosystem dynamics remains limited by the quality and scale of available data and models. Hence, in this doctoral thesis, we develop and evaluate several artificial intelligence methods, ranging from convolutional neural networks to large language models, to better map and understand European vascular plant species, essential biodiversity indicators, and terrestrial habitat types. Firstly, we assemble an extensive dataset that integrates millions of vascular plant species observations (both from citizen science and scientific experts data) combined with various environmental predictors at high-resolution (such as satellite images, climatic time series, and rasterized environmental variables). Building on this, we design a multimodal ensemble model to accurately predict species occurrences and derive biodiversity indicators for ecosystem health monitoring. We then focus on the core of the project: identification of habitats based on species assemblages, using deep learning and emphasizing interpretability to uncover ecological drivers. To capture the latent structure of plant communities, we introduce a novel use of large language models, fine-tuned to learn the syntax of co-occurring plants and thus able to detect missing species from incomplete surveys. Finally, we combine all these steps in a cascading pipeline capable of mapping species distributions, biodiversity indicators, and habitat types at high resolution across Europe, enabling previously infeasible large-scale ecological assessments by offering new tools for biodiversity monitoring, conservation planning, and land-use management. This work demonstrates the potential of modern machine learning approaches to address longstanding challenges in ecology. Using remote sensing, environmental variables, and species occurrence data, we built a pipeline offering a comprehensive view on ecosystems dynamics. By making all discoveries open source, we provide a scalable framework for future biodiversity assessment and conservation planning.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.