March 25, 2026 14:00 CEST
Live on Microsoft Teams.
On 25 March at 14:00 CEST, the ESA Φ-Lab Collaborative Innovation Network will host a new Φ-talk. Details are below.
Meet the speakers
Sébastien Lefèvre holds a Master of Science (1999), a PhD (2002), and a Habilitation (2009) in Computer Science. He has been a Full Professor at Université Bretagne Sud since 2010, promoted to the exceptional class by the National Council of Universities in 2023. He is also Adjunct Professor at UiT – The Arctic University of Norway and Visiting Professor at ESA Phi-lab. His work focuses on Artificial Intelligence for Earth and Environmental Observation. He founded the OBELIX team at IRISA and currently chairs the GeoData Science track of the EMJM Copernicus Master in Digital Earth. He co-founded the ECML-PKDD MACLEAN workshop series, chairs the AI4EO 2025 symposium, coordinates the UBS–JRC Doctoral Program on AI4EO, and holds the PANORAMIX chair within the SequoIA AI cluster (2025–2029). He serves on the scientific committee of IGN – the French Mapping Agency, and is active in ELLIS, IEEE GRSS, ISPRS, and IAPR. His research interests include image analysis and deep learning for remote sensing. As Earth Observation requires more than standard AI solutions, Prof. Lefèvre aims to conduct cutting-edge research in Artificial Intelligence to develop novel, efficient, scalable, and responsible solutions for high-impact tasks that leverage complex remote sensing data.
Pierre Adorni is currently a PhD candidate at IRISA, Université Bretagne Sud, France. He received his engineering degree in Computer Science and his M.Sc. in Machine Learning & Optimisation from UTC, Compiègne, France, in 2024. His research interests include deep learning, computer vision, representation learning, and foundation models for Earth observation. His current work focuses on improving the generalisation capabilities of machine learning models across diverse remote sensing datasets and tasks.
Talk abstract
Recent advances in foundation models have shown great promise in domains such as natural language processing and computer vision, and similar efforts are now emerging in the Earth Observation community. These models aim to generalise across tasks with limited supervision, reducing the need for training separate models for each task. However, current strategies, which largely focus on scaling model size and dataset volume, require prohibitive computational and data resources, limiting accessibility to only a few large institutions. Moreover, this paradigm of ever-larger models stands in stark contrast with the principles of sustainable and environmentally responsible AI, as it leads to immense carbon footprints and resource inefficiency. In this talk, we review some of our recent works that aim to lower the training cost of such models. We introduce an Ensemble-of-Specialists framework that decomposes the training process into lightweight, task-specific ConvNeXtV2 specialists that can be frozen and reused. This modular approach offers strong advantages in efficiency, interpretability, and extensibility. Moreover, it naturally supports federated training, pruning, and continuous specialist integration, making it particularly well-suited for collaborative and resource-constrained settings.