Summer Camp for Satellites

Seamus Geraty attended SatCamp in Boulder, Colorado from September 24-26, 2025. Here are his reflections on the experience.

This year’s SatCamp in Boulder was a fury of trail-side conversation, deep dives, and technical insight. A three-day “conference”, if it’s even fair to call it that, felt more like summer camp: running around with a bunch of your peers, talking shop and dreaming big against the majestic backdrop of the Boulder flatirons. This year’s theme of “Innovation and Disruption” set the stage, and many sessions focused on the quick pace at which our field is evolving alongside breakthroughs in generative AI.

One standout session for me was a panel on GeoAI and embeddings, featuring speakers from DevSeed, Legend, Planet, and the World Resources Institute. The focus was on how embeddings (compressed vector representations of complex geospatial data) along with foundational models are revolutionizing remote sensing. By transforming unstructured imagery into structured data, these algorithms allow machines to interpret satellite images with a newfound speed and precision. This can bring with it opportunity for scaling, automating and increasing accessibility for Earth observation.

There was palpable excitement around foundational models, which many felt are finally hitting their stride. Others noted we’re still just scratching the surface, with plenty of work ahead to enable wider adoption. But nearly everyone agreed: these tools aren’t a fad, they’re here to stay. Foundational models can bootstrap themselves with minimal training data, adapting to new areas of interest without starting from scratch. That’s a game-changer for labeling workflows, cost reduction, and analysis in data-sparse environments and last-mile use cases. We explored examples like Clay, an open-source Earth observation model built for accessibility and impact; AlphaEarth, DeepMind’s virtual satellite that stitches together multimodal EO data; and Global Nature Watch, a WRI-led platform that expands monitoring beyond forests to include wetlands, croplands, and carbon flux. The Earth Index from Earth Genome also stood out, its embeddings make the entire planet searchable, enabling rapid environmental intelligence from satellite imagery.

So what’s next? Panelists agreed that we need to see continued improvements in UI/UX design that opens these tools to a wider audience. Take fine-tuning for example, platforms now allow users to customize foundational models in simple, non-technical ways, accelerating adoption. The fAIr HOTOSM platform is a great example, integrating open-source models like RAMP. RAMP is designed to extract building footprints in low-resource settings, helping local teams generate high-quality maps with minimal data. fAIr, meanwhile, is an AI-assisted mapping service that lets users fine-tune RAMP for their own region, no coding required. It’s a powerful illustration of how thoughtful design and model customization are expanding access to geospatial AI.

That’s not to say that the last-mile challenges of applying these tools in these data-sparse environments have disappeared, there is still work to be done. But foundational models are making some large strides in lowering the barrier to entry. With smarter interfaces and more intuitive workflows, we are seeing a shift from elite research labs to grassroots deployment. SatCamp was a timely reminder that the future of GeoAI isn’t just a bunch of technical talk: its social, collaborative and deeply human, like a trail run with a bunch of newfound friends.