First On-Orbit Demonstration of a Geospatial Foundation Model

TL;DR

This work demonstrates that geospatial foundation models can be compressed and adapted for onboard EO processing, enabling a single pretrained backbone to power multiple tasks with lightweight task heads. Through dual-MAE distillation, a 16x smaller variant (256-MAE-D) preserves performance across five downstream tasks, despite notable domain shift between ground data and Kanyini imagery. Domain adaptation with pretrained task heads reduces label requirements to as little as 25% and improves transfer to the target domain, while validation on flight-representative hardware confirms feasible, energy-conscious onboard inference. An on-orbit demonstration on the IMAGIN-e platform establishes the first GeoFM deployment on an EO space payload, highlighting both the potential and practical challenges of autonomous AI in space missions.

Abstract

Geospatial foundation models (GeoFMs) promise broad generalisation capacity for Earth observation (EO) tasks, particularly under data-limited conditions. However, their large size poses a barrier to deployment on resource-constrained space hardware. To address this, we present compact variants of a Vision Transformer (ViT)-based GeoFM that preserve downstream task performance while enabling onboard execution. Evaluation across five downstream tasks and validation in two representative flight environments show that model compression and domain adaptation are critical to reducing size and resource demands while maintaining high performance under operational conditions. We further demonstrate reliable on-orbit inference with the IMAGIN-e payload aboard the International Space Station. These results establish a pathway from large GeoFMs to flight-ready, resource-efficient deployments, expanding the feasibility of onboard AI for EO missions.

Figures (4)

• Figure 1: Overview of the onboard GeoFM deployment pipeline and the satellite platforms used in this study. (a) A pretrained GeoFM is compressed, paired with task heads, and deployed for onboard inference. Downlinked imagery supports continual domain adaptation, with updated weights uplinked during operations. (b) Kanyini flight model and (c) its EO payload, HyperScout-2. (d) IMAGIN-e payload and (e) its mounting location on the ISS.
• Figure 2: Performance comparison between the original Prithvi-300M szwarcman2024prithvi and its 16$\times$ smaller variant, 256-MAE-D, across five downstream tasks. Results span cloud detection francis_alistair_2020_4172871(a, b), flood detection bonafilia2020sen1floods11(c), landslide detection ghorbanzadeh2022outcomeghorbanzadeh2022landslide4sense(d), and AGB estimation nascetti2023biomassters(e). 256-MAE-D delivers comparable performance across all tasks, showing that distillation preserves task generality while enabling onboard deployment.
• Figure 3: Performance of our GeoFM (256-MAE-D) on cloud and flood detection tasks, comparing held-out imagery from fine-tuning (source domain) with imagery from the Kanyini mission (target domain). Cloud detection (a, b) shows substantial declines across all metrics, whereas flood segmentation (c) exhibits modest gains.
• Figure 4: Effect of our GeoFM (256-MAE-D) on cloud and flood detection performance in the Kanyini domain under reduced labelled training data. Results are shown for tile-level cloud classification (a), cloud segmentation (b), and flood segmentation (c), with error bars denoting standard deviation across five random seeds. Baseline values in each subfigure indicate performance before domain adaptation, i.e., direct transfer from the source to the Kanyini domain.