Google Teams Unite on Earth AI: Boosting Geospatial Reasoning by 64% with Three Core Data Types

Earth AI Data System

Training relies on three specialized geospatial datasets. The imagery component aggregates three large‑scale remote‑sensing collections: RS‑Landmarks (18 M satellite/aerial images with high‑quality captions), RS‑WebLI (over 3 M open‑source images with scalability to billions), and RS‑Global (30 M images covering the globe at 0.1‑10 m resolution from 2003‑2022). These resources support visual‑language modeling, open‑vocabulary object detection, few‑shot learning, and backbone pre‑training for remote‑sensing research.

The population dynamics dataset merges built‑environment, natural‑factor, and human‑behavior information and applies graph neural networks to generate unified regional embeddings. Spatial coverage expands to 17 countries (including Australia, Brazil, India) with knowledge‑graph alignment for cross‑language and cross‑country pattern recognition. A monthly dynamic embedding series has been constructed from July 2023 onward. The label space spans health, socioeconomic, and environmental indicators and incorporates Yale PopHIVE county‑level COVID‑19 visit data and EU NUTS‑3 statistics.

Environmental data integrate weather, climate, and disaster sources, providing 240‑hour hourly forecasts, 10‑day daily forecasts, real‑time flood monitoring, and an experimental stochastic cyclone prediction system that generates 50 possible tracks up to 15 days in advance.

Earth AI Architecture

Three foundational modules (imagery, population, environment) are linked through a Gemini‑driven reasoning Agent. The Agent parses natural‑language or map‑based queries, decomposes them into sub‑tasks, dispatches the appropriate model tools, and synthesizes the results into coherent conclusions.

Model cooperation follows a “spatial alignment + representation integration” strategy: outputs from the AlphaEarth imagery model (terrain, climate) are mapped to common geographic units and merged with population embeddings to produce comprehensive regional portraits. Training proceeds in two stages—offline pre‑training on diverse geospatial signals to learn compact embeddings, followed by dynamic fine‑tuning for downstream tasks such as interpolation, extrapolation, super‑resolution, and nowcasting.

Benchmark Performance

Imagery foundation model achieves state‑of‑the‑art zero‑shot classification and text‑retrieval on public benchmarks (SigLIP2 / MaMMUT). Open‑vocabulary detection mAP reaches 31.83 % on DOTA and 29.39 % on DIOR; few‑shot training with only 30 samples per class lifts mAP above 53 %, surpassing existing methods. The pre‑trained ViT backbone improves average downstream performance by 14.93 % across 13 tasks and sets new records on FMOW classification and FLAIR segmentation.

Population dynamics model maintains stable R² when predicting missing variables for 20 % of regions and demonstrates strong cross‑country generalization. Monthly embeddings (from July 2023) reduce mean absolute error in COVID‑19 and influenza emergency‑visit forecasts, especially during peak winter periods; third‑party validation confirms robustness.

In multimodal fusion experiments, combining the population model with AlphaEarth raises R² by 11 % for FEMA disaster‑risk scoring and improves CDC health‑indicator predictions by 7 % (population only) and 43 % (AlphaEarth only). Integrated cyclone‑impact and disease‑risk forecasts cut baseline RMSE by 34 %.

The Gemini‑driven Agent was evaluated on a 100‑question benchmark (overall score 0.82) and a 10‑scenario crisis suite, outperforming Gemini 2.5 Pro by 64 % and Flash by 110 %. Iterative optimization yields consistently higher Likert scores on complex multi‑step geospatial reasoning tasks.

Technical Contributions and Related Work

By unifying large‑scale remote‑sensing, demographic, and environmental data under a foundation‑model and LLM‑driven Agent, the system lowers the expertise barrier for real‑time planetary analysis and enables actionable insights for climate response, disaster mitigation, and resource management.

Related academic work includes the EarthMind framework (University of Trento, TUM, Berlin Institute of Technology, INSAIT) and Stanford’s Marble 3‑D world‑generation model.

Earth AI: Unlocking Geospatial Insights with Foundation Models and Cross‑Modal Reasoning , arXiv 2510.18318, https://doi.org/10.48550/arXiv.2510.18318

EarthMind: Towards Multi‑Granular and Multi‑Sensor Earth Observation with Large Multimodal Models, https://doi.org/10.48550/arXiv.2506.01667