Google brings Gemini reasoning to satellite imagery

Assessing the aftermath of a major hurricane used to take a team of GIS experts several days, juggling satellite imagery, infrastructure layers, flood maps, and census records across clunky software.

Google wants to reduce that entire workflow to a single natural language prompt.

Google Research just unveiled Geospatial Reasoning, a new framework powered by generative AI and specialized remote sensing foundation models. To prove it is ready for prime time, Google signed up the biggest names in commercial Earth observation, including Airbus, Maxar, and Planet Labs, alongside ad giant WPP.

Giving Gemini a satellite-eye view

Analyzing satellite imagery is notoriously tricky because real-world spatial data is massive, messy, and hard to cross-reference with real-time conditions.

Google's new remote sensing models are built on proven computer vision architectures like SigLIP, MaMMUT, OWL-ViT, and masked autoencoders. Trained on high-resolution aerial and satellite shots tagged with text descriptions and bounding boxes, these models can handle zero-shot retrieval. You can simply ask for impassable roads or houses with rooftop solar panels, and the system flags them instantly.

Behind the scenes, Google is using Gemini as an orchestrator. When a user throws an open-ended crisis question at the system, Gemini spins up an agentic chain of reasoning.

In a hurricane demo, Gemini pulled high-res aerial imagery from the Civil Air Patrol, ran building damage models from Google's Open Buildings and SKAI, layered in NOAA flights, and cross-referenced the results against local housing prices, social vulnerability indices, and WeatherNext forecasts.

At the same time, Google expanded its Population Dynamics Foundation Model beyond the US to the UK, Australia, Japan, Canada, and Malawi, with WPP's Choreograph already plugging it into media planning.

Why it matters

We think Google's sharpest move here is treating geospatial intelligence not as a standalone computer vision feature, but as a modular toolset for autonomous agents.

For enterprise analysts, climate planners, and emergency responders, the barrier to complex spatial analysis has historically been specialized software and specialized headcount. Being able to run cross-dataset GIS queries through plain text directly inside Google Cloud will save critical hours when disasters hit.

For the Earth observation industry, having Airbus, Maxar, and Planet Labs on board validates a massive shift: nobody wants to buy raw satellite pixels anymore; they want structured answers. Google is positioning Gemini to be the operating system for planet-scale data.

That said, the stakes here are way higher than standard chatbot use cases. When an AI hallucination mislabels a flooded evacuation route or miscalculates structural damage after a storm, real lives are on the line. Building rock-solid guardrails around Gemini's multi-step spatial reasoning will be the real hurdle before public agencies fully trust it.