Google Earth AI: Unlocking geospatial insights with foundation models and cross-modal reasoning

Google has quietly been building a stack of AI models that watch the planet: satellite imagery analysis, population movement tracking, weather and flood forecasting. Individually, these have powered things like Maps updates and Search disaster alerts for years. What's new is that Google is now stitching them together with a Gemini-based orchestrator, called a geospatial reasoning agent, that can take a messy real-world question and actually work through it step by step instead of spitting out a single model's output.

The example Google keeps returning to is a good one: which specific neighborhoods are vulnerable to an incoming storm, and what infrastructure is at risk there. Answering that isn't a one-model job. The agent has to pull hurricane wind forecasts from an environment model, cross-reference county population data from Data Commons, grab official boundaries from BigQuery, run a spatial overlap, train a small model on the fly to rank postal codes by vulnerability, then hand a satellite image of the worst-hit zip code to an object-detection model to spot critical infrastructure. That's six distinct reasoning steps, chained automatically.

The underlying models themselves got meaningfully better too. The new Remote Sensing Foundations model improved text-based satellite image search by more than 16% on average and more than doubled baseline accuracy on zero-shot object detection in imagery it hadn't seen trained examples for. Population Dynamics Foundations, meanwhile, now has globally consistent embeddings across 17 countries updated monthly, and Oxford researchers found that folding those embeddings into a dengue fever forecasting model for Brazil pushed the 12-month prediction R² from 0.456 up to 0.656 — a serious jump for epidemiological forecasting, where marginal gains usually come in fractions of that size.

What's arguably more interesting than any single model is what happens when Google fuses them. Blending socio-economic embeddings from Population Dynamics with landscape data from AlphaEarth Foundations improved predictions on FEMA's National Risk Index by 11% on average across 20 hazard types, with tornado risk prediction jumping 25% and river flooding 17%. On Google's own benchmark, the full reasoning agent scored 0.82 accuracy on geospatial Q&A tasks, versus 0.50 for a bare Gemini 2.5 Pro agent and 0.39 for Flash — evidence that bolting specialized tools onto a general model matters more than just using a bigger general model.

The real test, though, is who actually uses this. Google name-checks a handful of early adopters: Bellwether, an Alphabet moonshot predicting storm damage to speed up insurance payouts; UN Global Pulse using it for post-disaster damage assessment; GiveDirectly routing cash aid to flood-threatened households before disaster hits. None of these are hypothetical use cases dreamed up for a blog post — they're organizations already leaning on the system for decisions with real consequences. Google is now opening access more broadly to enterprises and developers, which is the point where we'll find out whether this is genuinely useful infrastructure or just an impressive demo with a few sympathetic pilot partners.