Google's Geospatial Reasoning Agent Connects Gemini to Satellite and Climate Data

Google Research has introduced a Geospatial Reasoning agent that connects Gemini to Earth-specific foundation models, global datasets, and a dedicated planner. Satellite rasters, weather grids, and population layers are often too large and structured for an LLM context window. The agent creates an analytical plan, runs specialized tools in an external workspace, and synthesizes the results as maps, tables, and written reports with data attribution.

Preview access, two ways

Google is expanding experimental access through two interfaces described in its technical report.

Google also provides recipes for connecting the service to Gemini-based business workflows. Access requires an interest form, and the announcement does not specify pricing, quotas, service-level commitments, supported regions, or a production release date.

Four parts, one execution loop

• Earth-specific models and data. Gemini coordinates models including WeatherNext, Flood Forecasting, the Population Dynamics Foundation Model, AlphaEarth Foundations, and Remote Sensing Foundations. It can also query Google Maps Platform, Earth Engine, and Data Commons.
• Planning before execution. The agent decomposes a request into analytical steps, chooses fallback datasets when local coverage is sparse, and tracks source attribution. Large spatial layers remain in a dedicated workspace, limiting the amount of raw data passed through Gemini’s context window.
• Private data support. API users can supply CSV files, GeoJSON boundaries, and cloud tables, then join private locations or observations with Google’s global models and datasets.
• Automated spatial modeling. The Planetary Prediction Engine accepts local outcomes such as disease rates, crop yields, or grid outages. It selects relevant environmental and population signals, combines Population Dynamics Insights with Custom Satellite Embeddings, and trains and evaluates models for forecasting, spatial prediction, and downscaling coarse estimates to finer grids.

Specialized tools lift the scores

The report evaluates four configurations with Gemini 3.8 Flash and Gemini 3.1 Pro:

1. A standalone language model
2. The model with web search and maps grounding
3. The model with Earth AI tools and no upfront planning
4. The full agent with Earth AI tools and a dedicated planning stage

The evaluation covers two community benchmarks and two new Google benchmarks.

Google has published the development splits for GeoExpertQueries and GeoKTools on GitHub.

Adding Earth AI skills improved the reported results on GeoExpertQueries, GeoKTools, and CBGB. The gains exceeded those from changing the underlying Gemini variant, indicating that tool access had the larger effect on quantitative spatial tasks. UnivEARTH behaved differently: a standalone model scored 93.6%, which the authors attribute to the source articles and answers being available on the public web.

Upfront planning produced its clearest gains on GeoExpertQueries, where tasks require several dependent analytical steps. The planning stage reduced unsupported assumptions and discouraged shortcuts such as answering from web search when the task required spatial analysis.

Workflow scores leave an accuracy gap

GeoExpertQueries uses an automated grader to compare the agent’s process with an expert-authored reference workflow. It does not verify whether every final map, figure, or conclusion is correct. A system can therefore follow a plausible sequence of steps and still produce a faulty result, while an alternative valid method may diverge from the reference process.

Google positions the output as a first draft for expert review. High-stakes deployments in public health, disaster response, infrastructure, and resource allocation still require checks against source data, domain constraints, and known outcomes.

Pilots span outbreaks, storms, and solar

Google tested the agent with social-impact and commercial organizations in a closed pilot and reported four operational examples:

• Ebola response in the Democratic Republic of the Congo. WHO AFRO cross-referenced satellite-derived forest loss, mining locations, web data, and population estimates to map remote mining corridors. Google says the analysis identified 48 exposed settlements and more than 45,500 people at risk within minutes.
• Cyclone preparedness. People in Need analyzed 14 historical cyclone paths to identify villages with high exposure, reducing work that previously required days of manual GIS analysis.
• Crisis scoping. GiveDirectly uploaded custom boundaries to estimate population exposure across wind-speed zones during Tropical Cyclone Maila and assess tornado damage in Mississippi. Google says the shared workspace reduced the process from days to minutes.
• Site selection. Milken Institute teams combined more than 1,500 candidate buildings with the Google Maps Platform Solar API and census data. The screening considered rooftop solar potential, mobility, transit access, climate risks, and federal investment eligibility.

The published pilot summaries provide speed and scale figures while omitting error rates, comparison methods, and independent validation.

The design travels beyond maps

Google’s architecture separates language reasoning, domain execution, and data storage. A planner decomposes the request, specialized tools perform the calculations, and an external workspace holds intermediate data before Gemini prepares the final response. This structure avoids passing entire datasets through the model and gives developers clearer places to add validation, provenance tracking, retries, and access controls.

The same design can support other domains where source data is too large or structured for direct prompting, including observability logs, genomic datasets, financial records, and large code repositories. Its effectiveness depends on the quality of the available tools, the planner’s ability to select them, and the checks applied to their outputs.

A separate public-health study cited by Google reported that combining population and satellite signals improved health-outcome predictions by 12% over demographic baselines and supported mapping at eight times finer spatial resolution. Those results cover a specific research setting, but they illustrate the potential benefit of combining domain models through an orchestrated workflow.

Production questions remain open

• Cost and capacity: Google has not published pricing, quotas, latency targets, or service guarantees.
• Data governance: Teams handling private locations or health data will need terms covering retention, training use, residency, and access controls.
• Failure handling: Developers need documented tool errors, retry behavior, timeouts, and observability for multi-step jobs.
• Evaluation: Each deployment needs representative test cases with verified outputs, including sparse-data regions and conflicting sources.
• Review: Human approval and source-level inspection remain necessary for consequential decisions.