Following the massive success of large language models, geospatial foundation models have also gained momentum in the GIS community. On July 12, Esri officially introduced the next-generation Geospatial Foundation Models in ArcGIS at its annual User Conference, integrating foundational geospatial AI capabilities directly into ArcGIS workflows. Whether you are working on remote sensing interpretation, site selection analysis, similar feature retrieval, or keeping an eye on Google’s AlphaEarth satellite embedding dataset, these Geospatial Foundation Models deserve your attention. Today I will give a brief introduction.

The newly added capabilities are divided into three complementary categories: Location Encoder Models, Geospatial Vision Language Model (GeoVLM), and Remote Sensing Foundation Models.

Understanding Embedding

Before diving into the models, we need to understand a foundational concept: Embedding. In ArcGIS, embeddings are treated as a new type of GIS data that can be stored, analyzed, and overlaid with traditional vector, raster, and tabular data. An embedding can be thought of as a compact numerical vector that summarizes the key characteristics of an image, a location, or other geographic features—essentially a compressed representation of geographic information.

Location Encoder Models

Location Encoder Models upgrade a location from simple latitude and longitude to a set of embedding vectors that capture the local natural environment, built-up area morphology, socio-economic context, and more. These embeddings can be used for similarity search, clustering, predictive modeling, site selection, change detection, and other tasks. Esri has released two complementary versions. One generates natural environment embeddings, trained on Sentinel-2 imagery, and can produce embeddings for any location through DLPK on Living Atlas. The other produces statistical data embeddings, currently trained on thousands of variables from U.S. Census data, the American Community Survey (ACS), housing, and environmental statistics.

Geospatial Vision Language Model (GeoVLM)

GeoVLM (Geospatial Vision Language Model) brings multimodal AI to earth observation. Using natural language prompts, GeoVLM supports object detection, pixel classification and segmentation, image captioning, object counting, visual question answering, and image or region classification. The model is trained on millions of image-text pairs covering multiple geographic regions and is optimized specifically for remote sensing imagery rather than everyday photos. It is also published through Living Atlas as a DLPK.

Remote Sensing Foundation Models: TerraMind, Prithvi, Clay, and More

Remote Sensing Foundation Models represent another major breakthrough in geospatial AI. Traditional computer vision models are pre-trained on everyday photographs, but remote sensing foundation models are trained directly on satellite and aerial remote sensing data, such as Sentinel, Landsat, and NAIP imagery. As a result, they provide much stronger initialization for most earth observation tasks, require less labeled training data, and deliver higher analysis accuracy.

Esri is both developing its own next-generation remote sensing foundation models and integrating several popular open-source remote sensing foundation models into ArcGIS. Users can generate image embeddings and fine-tune these models using ArcGIS Pro.

Currently supported models include:

  • TerraMind
  • Prithvi EO 2.0
  • Clay
  • DOFA
  • DINO

Summary

The release of Geospatial Foundation Models by Esri marks a significant transformation in the integration of GIS and AI. Organizations no longer need to develop separate models for each business task; they can directly leverage foundation models with built-in geographic understanding to adapt various GIS analysis workflows. Esri is not simply stacking models—it is embedding AI capabilities as a universal foundation within ArcGIS.

In my opinion, it will be worth watching whether Living Atlas models and statistical data can achieve global coverage (or open a path for custom training integration), as well as how Agent-based spatial analysis works when combined with ArcGIS MCP. Additionally, the current features are still in early Beta, but they serve as an excellent reference model for GIS professionals.

References