First On-Orbit Demonstration of a Geospatial Foundation Model

First On-Orbit Demonstration of a Geospatial Foundation Model Andrew Du1*, Roberto Del Prete3, Alejandro Mousist4, Nick Manser2, Fabrice Marre2, Andrew Barton2, Carl Seubert2, Gabriele Meoni3, Tat-Jun Chin1 1*AI for Space Group, The University of Adelaide, Adelaide, 5000, South Australia, Australia. 2SmartSat CRC, Adelaide, 5000, South Australia, Australia. 3Φ-lab, European Space Agency, Frascati, 00044, Italy. 4Thales Alenia Space, Cannes, 06150, France. *Corresponding author(s). E-mail(s): andrew.du@adelaide.edu.au; Contributing authors: Roberto.DelPrete@esa.int; Alejandro.Mousist@thalesaleniaspace.com; nick.manser@smartsatcrc.com; fabrice.marre@smartsatcrc.com; andrew.barton@smartsatcrc.com; carl.seubert@smartsatcrc.com; Gabriele.Meoni@esa.int; tat-jun.chin@adelaide.edu.au; 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 Trans- former (ViT)-based GeoFM that preserve downstream task performance while enabling onboard execution. Evaluation across five downstream tasks and valida- tion in two representative flight environments show that model compression and domain adaptation are critical to reducing size and resource demands while main- taining 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. 1 arXiv:2512.01181v1 [cs.LG] 1 Dec 2025 Keywords: Earth observation, satellite, geospatial foundation model, machine learning, model compression, domain adaptation 1 Introduction Understanding Earth’s dynamic systems through satellite imagery is critical for addressing a wide range of environmental and societal challenges, such as monitoring climate change [1, 2], managing natural resources [3, 4], supporting sustainable infras- tructure [5–7], and enabling timely responses to natural disasters [8, 9]. To meet these needs, an increasing number of Earth observation (EO) satellites have been launched or are planned for launch, equipped with ever-advancing imagers capable of captur- ing high-resolution multispectral and hyperspectral imagery across a broad swath of the electromagnetic spectrum. However, the primary challenge in EO has shifted from acquiring data to efficiently analysing and extracting actionable insights from the vast volumes collected, particularly in bandwidth-limited or real-time scenarios. As a result, there is growing interest in deploying machine learning (ML) techniques, particularly deep neural networks (DNNs), directly onboard satellites to enable more advanced processing and analysis in orbit. Recent advances in space-qualified hardware acceler- ators, such as Intel’s Myriad Vision Processing Unit (VPU) series [10–13], Ubotica’s XE platforms [14–16], and NVIDIA’s Jetson modules [17–19], have begun to make this feasible, opening the door to more capable and intelligent EO systems. To date, several EO missions have demonstrated or plan to demonstrate the fea- sibility of deploying ML onboard satellites. Φ-sat-1 [10, 20] marked the first use of a DNN on a space-qualified artificial intelligence (AI) chip (i.e., Intel Movidius Myriad 2), executing a convolutional neural network (CNN) in orbit to discard cloudy images. Building on this, Φsat-2 [21, 22] plans to support multiple onboard applications, includ- ing cloud detection, marine vessel tracking, image compression, anomaly detection, and wildfire detection, though these have yet to be demonstrated. ION-SCV 003 [12] showed that models can be iteratively updated by capturing images onboard, down- linking them for labelling, retraining on the ground, and uplinking the revised weights. Its successor, ION-SCV 004 [13], and SONATE-2 [19] extended this capability by sup- porting onboard training on preloaded datasets acquired from previous missions. More recently, CogniSAT-6 [16, 23] became the first EO satellite to integrate onboard AI processing, autonomous task scheduling, and near real-time insight delivery via inter- satellite links (ISL). It has also executed over 20 onboard applications including flood detection, land cover classification, and volcano monitoring [24, 25]. Separately, ISS Mounted Accessible Global Imaging Nod-e (IMAGIN-e) [26, 27] has enabled AI pro- cessing capabilities aboard the International Space Station (ISS), serving as a testbed for edge computing in space without the need for a dedicated satellite. While these missions demonstrate meaningful progress, they largely rely on lightweight, task-specific models—typically CNNs—due to three practical constraints: limited onboard compute, memory

This content is AI-processed based on open access ArXiv data.