Google has unveiled a major leap forward in geospatial intelligence through the launch of new geospatial foundation models and a complementary research initiative called Geospatial Reasoning. These innovations integrate generative AI with geospatial data to accelerate problem solving in critical areas such as disaster response, public health, climate resilience, and commercial operations.
Geospatial data—information linked to physical locations—has long powered familiar Google products like Maps, Earth, and Search. Now, Google is applying its decades of experience in organizing and delivering geospatial information to build AI systems that transform this data into actionable insights. This effort tackles longstanding challenges in the field, such as data complexity, inconsistency across sources, and the difficulty of gathering labeled observations. Most recent AI models have not been optimized for geospatial use cases, creating an urgent need for tailored solutions.
At the core of this initiative are two new pre-trained models: the Population Dynamics Foundation Model (PDFM) and a mobility-based foundation model. The PDFM captures the relationship between population behavior and the local environment. It has already been tested by over 200 organizations across the U.S., and the dataset is now expanding to include regions such as the UK, Australia, Japan, Canada, and Malawi.
Complementing these tools, Google has introduced advanced remote sensing models. These geospatial foundation models are trained on high-resolution satellite and aerial imagery paired with text annotations and bounding boxes. Built on architectures like masked autoencoders, SigLIP, MaMMUT, and OWL-ViT, they produce rich visual embeddings and support tasks like road and building mapping, infrastructure detection, and disaster impact analysis. The models also support zero-shot classification, enabling users to find images based on plain-language queries such as “residential buildings with solar panels” or “impassable roads.”
Evaluations show that these models achieve state-of-the-art performance across remote sensing benchmarks involving classification, segmentation, and object detection. They have already been deployed internally at Google for applications such as urban and agricultural mapping and emergency response, consistently improving performance on task-specific metrics.
To tie these capabilities together, Google has launched Geospatial Reasoning—a framework that allows developers and analysts to create agentic workflows using large language models like Gemini. This system enables users to interact with complex datasets through natural language, performing sophisticated geospatial queries that blend public, proprietary, and real-time data. In one use case, a crisis manager could visualize satellite imagery before and after a hurricane, detect building damage using AI, query Gemini for damage assessments by neighborhood, and receive recommendations for relief prioritization based on social vulnerability indices.
The full system includes a Python front-end application, a LangGraph-powered backend, and tools that access Earth Engine, BigQuery, Maps Platform, and other services. The models are currently being tested by partners like Airbus, Maxar, and Planet Labs.
With the launch of these geospatial foundation models, Google is creating a unified AI-driven platform for understanding the physical world. By grounding generative AI in spatial data, these tools offer transformative potential for governments, researchers, and businesses alike. As Google continues to expand access, geospatial foundation models will play a foundational role in addressing global challenges with speed, accuracy, and scale.