Google Research is expanding early access to a geospatial agent that combines Gemini with weather, flood, population, satellite and mapping tools. The experimental system can plan a multi-step spatial analysis, execute it in a dedicated workspace and show the datasets and assumptions behind its answer.
Organizations can add CSV files, custom map boundaries or cloud tables to public Earth AI data. A related Planetary Prediction Engine can select environmental and population signals and train local models for forecasting or downscaling. Google says combining satellite and population signals improved health-outcome predictions by 12 percent over demographic baselines in its research and produced maps at eight times finer resolution.
During an Ebola outbreak in the Democratic Republic of the Congo, a World Health Organization regional team used the agent to combine mining, forest-loss and population data. Google reports that it identified 48 exposed settlements and more than 45,500 people at risk in minutes rather than weeks. Humanitarian group People in Need similarly analyzed 14 historic cyclone paths to locate high-exposure villages.
Google’s benchmarks show that specialized spatial tools and upfront planning improved multi-step work over a standalone language model. One benchmark, however, uses an automated grader to compare workflows rather than verify final answers. Google therefore describes results as first drafts for expert review. Access is expanding through Ask Google Earth and an experimental agent-to-agent service, not a general production release.