2nd ESA-NASA Workshop on AI Foundation Model for Earth Observation

Building on the success of the first ESA/NASA Workshop, hosted at ESA-ESRIN, the second Workshop in the series will be held May 19–22, 2026, hosted by NASA MSFC in Huntsville, Alabama.

The first workshop demonstrated growing alignment between the EO and AI-FM communities, with an explicit call for stronger coordination to reduce redundant efforts, minimize silos and for ESA/NASA to act as catalysts for sustained cross-domain collaboration. A clear emphasis emerged on moving past rapid prototyping toward operational deployment, including designing FMs to be accountable, scientifically validated, interoperable, computationally efficient, and integrated into downstream systems (Digital Twins, dashboards, early warning, and edge/onboard use). The community highlighted reproducibility and transparent benchmarking as prerequisites for trust and adoption, calling for evaluation frameworks that reflect scientific rigor and operational constraints to help users navigate a fast-moving FM landscape.

Furthermore, the recommendations from the first workshop highlighted the need for continued refining benchmarks to better capture real EO conditions (noise/complexity), and expand evaluation to include model behavior, uncertainty and efficiency, and shifting benchmark definition from model builders to real end users. The workshop report asked to prioritize scalable methods (parameter-efficient adaptation, embedding-based workflows, neural compression), integrate domain knowledge to ensure physical plausibility/interpretability, and assess environmental impact across the FM lifecycle (training through inference).

Ultimately, the goal is to develop more highly dimensional, physically grounded models (multimodality/sensor, multitemporal, multiresolution), revisit pretraining/pretext tasks, integrate climate/weather principles, and include uncertainty quantification and evolving beyond static pipelines toward modular agentic AI that can dynamically reason and act on EO data (especially when paired with FM models for EO).

Read More