Putting Earth AI to work: Geospatial agents for our planet's biggest challenges
Google Research introduces a Geospatial Reasoning agent that combines Earth AI foundation models with upfront planning to solve complex spatial problems. Early tests with partners like WHO AFRO demonstrate rapid risk mapping for disease outbreaks, while new benchmarks validate the system's ability to handle multi-step workflows beyond standard LLM capabilities.
Tags
Quick summary
Google Research introduces a Geospatial Reasoning agent that combines Earth AI foundation models with upfront planning to solve complex spatial problems. Early tests with partners like WHO AFRO demonstrate rapid risk mapping for disease outbreaks, while new benchmarks validate the system's ability to handle multi-step workflows beyond standard LLM capabilities.
Putting Earth AI to work: Geospatial agents for our planet's biggest challenges
General-purpose large language models (LLMs) and standard coding tools often struggle with real-world spatial problems. Planetary datasets are massive, featuring specialized map and satellite formats that require careful multi-step planning to combine accurately. To address these complexities, Google Research has evolved a Geospatial Reasoning agent designed to carry out complex spatial analysis and planetary prediction that standalone LLMs cannot achieve.
This system connects the Gemini model to the capabilities of Earth AI's foundation models, including WeatherNext, Flood Forecasting, the Population Dynamics Foundation Model (PDFM), AlphaEarth Foundations, and Remote Sensing Foundations. It also integrates information from Google Maps Platform, Google Earth Engine, Data Commons, and Google Dataset Search. By packaging these capabilities as modular skills, the agent calls specific tools only when a user question requires them, ensuring efficiency and accuracy.
Architectural Evolution and Modular Skills
The Geospatial Reasoning agent represents a significant evolution from early concepts into an experimental system built for multi-step reasoning across planet-scale data. A core component of this architecture is the ability to perform upfront planning. Before executing complex operations, the agent builds a step-by-step analytical plan. It selects appropriate fallback datasets when local coverage is sparse and processes large map layers in a dedicated workspace.
This approach prevents the agent's context from being overwhelmed by massive amounts of spatial data while preserving accuracy and tracking data attribution. Every response generated by the agent includes the plan and the data sources used, allowing experts to check assumptions or substitute datasets as needed.
Furthermore, the system supports "Bring Your Own Data" workflows. Organizations can input their own datasets via the Geospatial Reasoning agent API as CSV spreadsheets, custom map boundaries in GeoJSON format, or cloud tables. Users can then ask the agent to join their private locations with Earth AI's global models and catalogs, facilitating a hybrid analysis of local and planetary data.
From Reasoning to Planetary Prediction
Beyond answering complex analytical questions, the agent powers the recently announced Planetary Prediction Engine. Given a user's local outcome data—such as regional disease rates, crop yields, or grid outages—the agent automatically selects relevant population and environmental signals. It utilizes Population Dynamics Insights and the annual Satellite Embedding dataset (built on PDFM and AlphaEarth Foundations, respectively) to train validated machine learning models.
These models are used for spatial prediction, super-resolution downscaling, and forecasting. In recent public health research, combining these population and satellite signals improved predictions of health outcomes by 12% over demographic baselines and enabled mapping at 8X finer spatial resolution. The agent responds to a user's question by first planning a multi-step workflow, selecting key Earth AI tools and datasets, and finally executing each stage in a workspace outside the context window. The final answer incorporates multiple Earth AI capabilities into a text, tabular, and map-based report.
Benchmarking Geospatial Reasoning
To measure the impact of domain-specific tools and upfront planning, Google Research conducted a technical evaluation comparing four agent configurations across Gemini 3.8 Flash and Gemini 3.1 Pro. The configurations ranged from a standalone LLM to agents progressively augmented with skills and orchestration optimized for geospatial reasoning.
The evaluation spanned four benchmarks:
- GeoKTools: Contains templated queries with verifiable answers, each requiring the correct selection and use of 1 to 4 distinct geospatial tools.
- GeoExpertQueries: Introduces complex, real-world queries with expert-authored golden workflows that capture analysis best practices in domains like crisis resilience, site selection, and public health.
- CBGB: An externally developed benchmark for geospatial code generation.
- UnivEARTH: An externally developed benchmark for question answering.
Two key takeaways emerged from the evaluation. First, domain-specific skills are essential for spatial reasoning. An agent equipped only with basic tools (web search, maps grounding, and code execution) struggles on quantitative spatial tasks. Equipping the agent with Earth AI skills drives a significant jump in performance across GeoExpertQueries, GeoKTools, and CBGB. The only exception was UnivEARTH, where even a standalone LLM scored 93.6%, likely because the source NASA articles and answers were available on the public web.
Second, upfront planning unlocks open-ended, real-world workflows. While Earth AI skills alone were sufficient for narrower tasks like those in CBGB and GeoKTools, tackling open-ended problems in GeoExpertQueries required upfront planning. This planning stage keeps the agent on track across multi-step analyses and prevents it from making unreliable assumptions or taking shortcuts, such as using web search instead of spatial tools.
The technical report notes that GeoExpertQueries scores come from an autorater that compares the agent's performance to expert workflows defining the steps an analysis should take, rather than just the correct final answer. Therefore, the agent's output is considered a strong first draft for experts to review and build upon, not a final answer.
Real-World Applications and Early Access
Following promising tests with partners, Google Research is expanding experimental agent access via Ask Google Earth and an Agent-to-Agent (A2A) service. Early access is being offered to an experimental deep research update for Ask Google Earth, as well as an agent-to-agent API for developers building custom workflows, including integration recipes for connecting to Gemini.
In real-world testing, a small group of social impact and commercial partner organizations, including the World Health Organization (WHO) Regional Office for Africa, People in Need, GiveDirectly, the Milken Institute, and WPP, tested the agent in support of their own workflows.
For instance, at the start of the 2026 Ebola outbreak in the Democratic Republic of the Congo (DRC), WHO AFRO's Emergency Preparedness and Response Hub in Dakar used the Geospatial Reasoning agent to map remote mining corridors where exposure risk and human mobility were high. By cross-referencing forest loss from satellite images, mining sites from maps and web data, and population data, the team pinpointed 48 exposed settlements and located more than 45,500 at-risk people in minutes—a process that normally would have taken weeks. This allowed local responders to proactively deploy mobile laboratories and coordinate border surveillance.
Simultaneously, the Epidemic Modeling and Intelligence Unit (UMIE) at the DRC's National Institute for Biomedical Research (INRB) used the Planetary Prediction Engine to estimate the risk of Ebola spreading into uninfected zones. By combining mobility flows with historical case trends and Earth AI models, these weekly insights provided coordinators with critical planning time before cases arrived. As noted in the research, mobile populations like miners move faster than traditional paper surveillance can track, creating dangerous blind spots. Earth AI fills those gaps with satellite and mobility signals, moving the response from reactively chasing a virus to proactive prevention.
Conclusion
The Geospatial Reasoning agent demonstrates how combining domain-specific foundation models, spatial tools, and upfront planning enables complex spatial analysis and planetary prediction. By distinguishing between pretraining capabilities and fine-tuned workflows, and by leveraging a dedicated spatial workspace, the system overcomes the limitations of general-purpose LLMs when facing planetary-scale data. As Google Research expands early access and releases new benchmarks for the community, the potential for AI to address our planet's biggest challenges continues to grow.
Source: https://research.google/blog/putting-earth-ai-to-work-geospatial-agents-for-our-planets-biggest-challenges
