Answers are grounded in this paper and cite their sources.
Geospatial foundation models (GeoFMs) offer broad generalisation across Earth observation (EO) tasks, particularly under data-limited conditions. However, their large model size and computational requirements remain major barriers to deployment on resource-constrained space hardware. To address this challenge, we develop compact variants of a GeoFM designed to preserve downstream task performance while enabling onboard execution. The models are evaluated across five downstream EO tasks and validated on two representative flight-computing environments. The results show that compression substantially reduces model size (1157.03 MB to 73.30 MB; approximately 16×) and enables execution on resource-constrained hardware (FP32-to-FP16 conversion for the Myriad-2 resulted in differences of no more than 0.21 percentage points across all reported task-level performance metrics), while task-head adaptation improves performance on mission-domain imagery under limited supervision. We further demonstrate on-orbit execution of the compact GeoFM using the IMAGIN-e payload aboard the International Space Station, where inference was performed on a preloaded curated dataset under orbital computing conditions. These findings establish a practical pathway for translating large pretrained GeoFMs into resource-efficient onboard implementations and expand the feasibility of onboard artificial intelligence for future EO missions.
No takes yet. Share an insight, caveat, or question.
Du et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: