Google is expanding access to Earth AI — a powerful geospatial system built on large language model, Gemini, to help cities, businesses and nonprofits respond to real-world crises faster. The system is built on three specialized AI model families: one for interpreting satellite imagery through natural language, another for understanding human population patterns, and a third for predicting weather, floods, and cyclones. The true innovation is a Gemini-powered reasoning agent that acts as a "geospatial brain." This agent can deconstruct complex, real-world questions—like identifying regions with high flood risk and significant vulnerable populations—and automatically orchestrate the correct models and data sources to provide a coherent, actionable answer.
Google is unveiling a powerful new geospatial intelligence system called Earth AI, a comprehensive approach built on its large language model Gemini. According to a detailed technical paper, Earth AI integrates decades of siloed data—including satellite imagery, population dynamics, and environmental signals—into a unified, AI-driven framework. The system is built upon a family of specialized foundation models that work in concert. Its imagery models can understand satellite photos using natural language, allowing users to ask complex questions about visual data. Its population models capture the dynamic patterns of human life by fusing anonymized data on behavior and mobility. Simultaneously, its environment models provide state-of-the-art forecasting for weather, floods, and cyclones.
The true breakthrough lies in the Gemini-powered reasoning agent that orchestrates these components. This intelligent agent acts as a geospatial brain, capable of deconstructing a complex query such as "Identify counties with high elderly populations in the hurricane's path." It then plans and executes a multi-step analysis, delegating tasks to the appropriate specialist models and synthesizing the results into a coherent, actionable answer. This process automates workflows that would typically require a team of expert analysts.
Rigorous evaluation underscores the system's practical utility. Research demonstrates that combining the different foundation models unlocks synergistic insights, leading to significantly superior predictive capabilities. In one application, predicting community risk scores for natural disasters was 11% more accurate using combined models. In a collaboration with the World Health Organization, Earth AI was used to forecast cholera outbreaks in the Democratic Republic of Congo, achieving a 34% reduction in forecast error by integrating population and weather data. Furthermore, in a retrospective analysis of Hurricane Ian, a partner used the system to predict the number of damaged buildings three days before landfall with an error of only 3%.
By creating an interoperable ecosystem of models orchestrated by advanced AI, Earth AI bridges the critical gap between vast planetary data and actionable understanding. This technology promises to democratize sophisticated geospatial analysis, enabling decision-makers to respond more effectively to the urgent challenges of a changing planet.
Sources:
Google Research
https://arxiv.org/abs/2510.18318v1 .
Provided by the IKCEST Disaster Risk Reduction Knowledge Service System
Comment list ( 0 )
