Artificial intelligence is reshaping how we understand our planet. Following the success of large language models, a new generation of geospatial foundation models is emerging that learns directly from maps, satellite imagery, demographic information, and other geographic data.
Unlike traditional AI models that are trained for a single task, geospatial foundation models learn rich, reusable representations of geographic information. Once trained, they can be adapted to many GIS workflows – from similarity search and predictive modeling to feature extraction and natural language interaction with imagery – with far less training data and effort.
At this year’s Esri User Conference, we are expanding ArcGIS with a new generation of geospatial foundation models that bring these capabilities to GIS professionals through familiar ArcGIS workflows.
At the core of many of these models are embeddings – compact numerical representations that capture the essential characteristics of an image, location, or other geographic feature. Similar places produce similar embeddings, enabling workflows such as similarity search, clustering, prediction, and retrieval. In ArcGIS, embeddings become first-class GIS data that can be stored, analyzed, and combined with traditional spatial information.
Geospatial foundation models require more than advances in AI. They require deep understanding of geographic data and the workflows used to analyze it. By combining decades of geospatial expertise, authoritative Esri datasets, and the flexible deployment patterns in ArcGIS, Esri is helping shape the next generation of AI for location intelligence and Earth observation.
The new capabilities in ArcGIS span three complementary areas: Location Encoder models, Geospatial Vision Language Models, and Remote Sensing Foundation Models.
Location Encoder Models
Location encoder models learn representations of places rather than individual images or features. Instead of treating a location simply as a pair of coordinates, they learn to capture the geographic, environmental, and socioeconomic characteristics that define each place.
The embeddings produced by these models provide a compact numerical representation of location that can be used across many GIS workflows, including similarity search, clustering, predictive modeling, interpolation, and site selection.
Esri has developed two complementary location encoder models.
Global Location Encoder
The Global Location Encoder (Sentinel-2) model learns from globally available Sentinel-2 imagery to produce embeddings that capture the characteristics of locations around the world.
By learning directly from satellite imagery, the model captures patterns describing both the natural and built environment without requiring manually engineered variables.
The resulting embeddings can support a variety of downstream workflows, such as similarity search, location retrieval, clustering, predictive modeling, and change detection, either directly or in combination with users’ own geospatial data.
The Global Location Encoder (Sentinel-2) is available through ArcGIS Living Atlas as a Deep Learning Package (DLPK), allowing users to generate embeddings for locations virtually anywhere on Earth.
Geodemographic Foundation Model
While the Global Location Encoder learns from satellite imagery, Esri’s Geodemographic Foundation Model learns from rich geodemographic information.
The model has been trained using thousands of authoritative demographic, socioeconomic, housing, and environmental variables, including data from the U.S. Census, the American Community Survey, housing datasets, and environmental sources.
Instead of working directly with thousands of input variables, users obtain compact embeddings that preserve the underlying geographic relationships while making downstream analysis simpler.
These geodemographic embeddings can be applied across a wide range of geospatial workflows, from similarity search and clustering to market analysis, site selection, spatial interpolation, and predictive modeling where demographic, socioeconomic, or environmental context plays an important role.
Evaluations across numerous predictive tasks have shown that these embeddings consistently improve predictive performance when demographic context is an important driver, particularly when combined with traditional explanatory variables.
The USA Geodemographic Embeddings, produced by this model, are being released as a beta feature layer through ArcGIS Living Atlas.
Geospatial Vision Language Model (GeoVLM)
Esri has also developed GeoVLM, a Geospatial Vision Language Model that brings the power of multimodal AI to Earth observation.
GeoVLM brings the power of multimodal AI to Earth observation by connecting satellite imagery with natural language. Instead of building separate models for every remote sensing task, users can interact with imagery using prompts.
GeoVLM can support a wide range of remote sensing tasks through natural language prompting, including object detection, pixel classification and segmentation, image captioning, object counting, visual question answering, and image or region classification.
The model has been trained on millions of image-text pairs spanning multiple geographic regions using both Esri-generated and open datasets. Because it is designed specifically for Earth observation, it understands remote sensing imagery rather than everyday photographs.
Geospatial Vision Language Model is being released through ArcGIS Living Atlas as a Deep Learning Package (DLPK), enabling organizations to incorporate multimodal AI into ArcGIS workflows while keeping their imagery and data within their own infrastructure.
Remote Sensing Foundation Models
Remote sensing foundation models represent another major advancement in geospatial AI.
Unlike traditional computer vision models pretrained on everyday photographs, these models are trained directly on satellite and aerial imagery from sensors such as Sentinel, Landsat, and NAIP. As a result, they provide stronger starting points for many Earth observation tasks, often requiring less labeled training data while delivering improved accuracy.
Alongside developing its own next-generation remote sensing foundation models, Esri has integrated ArcGIS with several leading open-source remote sensing foundation models. Users can leverage these models to generate embeddings for imagery and, where supported, fine-tune them for downstream geospatial deep learning tasks using ArcGIS Pro and the arcgis.learn API.
Currently supported models include:
- TerraMind
- Prithvi EO 2.0
- Clay
- DOFA
- DINO
These integrations make it easy for ArcGIS users to take advantage of the latest advances in Earth observation AI without leaving familiar GIS workflows.
In our next blog post, we’ll take a deeper look at these remote sensing foundation models, how they differ, and how they can be used for embedding generation and fine-tuning in ArcGIS.
Looking Ahead
Foundation models represent a significant shift in geospatial AI. Rather than building separate models for every application, organizations can begin with models that already understand geographic information and adapt them to a wide variety of GIS workflows.
With the introduction of the Global Location Encoder, the Geodemographic Foundation Model, GeoVLM, and support for leading remote sensing foundation models, ArcGIS provides a comprehensive platform for applying the latest advances in geospatial AI. Whether the goal is understanding places, analyzing imagery, or building predictive models, these new capabilities make foundation models more accessible to the broader GIS community.
Acknowledgments
We appreciate AWS for providing the cloud infrastructure that helped accelerate the training of these models
Article Discussion: