Fraunhofer Institute for Computer Graphics Research IGD
Fraunhofer Institute for Computer Graphics Research IGD

Geospatial Foundation Models for High Resolution Geospatial Data

Prediction of an orthophoto with a spatial resolution of 50cm using a fine-tuned Clay model
Prediction of an orthophoto with a spatial resolution of 50cm using a fine-tuned Clay model
Prediction of Dortmund orthophotos using the GeoNRW dataset and a Mask2Former model with a Vision Transformer backbone, pre-trained on Baltic Brain using digital orthophotos (DOPs) from Germany.
Prediction of Dortmund orthophotos using the GeoNRW dataset and a Mask2Former model with a Vision Transformer backbone, pre-trained on Baltic Brain using digital orthophotos (DOPs) from Germany.

Using AI for Geospatial Data – Even with Limited Training Data

Aerial imagery and other geospatial data provide valuable information for urban development, infrastructure, and monitoring. However, their automated analysis using AI often requires extensive, annotated training data. At the same time, transferring existing models to new regions or significantly higher-resolution data presents unique challenges.

Geospatial foundation models can open up new possibilities here: these pre-trained AI models can be adapted for specific analysis tasks and reduce the need for application-specific training data.

Fraunhofer IGD is investigating which foundation models are suitable for high-resolution geospatial data, how they can be efficiently adapted, and when established AI methods provide the better solution.

Foundation Models for High-Resolution Aerial Imagery

Many available geospatial foundation models have previously been trained primarily on medium- to low-resolution satellite data, including approaches from IBM and Google. For applications in urban development, infrastructure, and public tasks, however, high-resolution aerial imagery and digital orthophotos are frequently available.

Fraunhofer IGD is therefore investigating how such geospatial foundation models, as well as vision foundation models like Meta's Segment Anything Model (SAM), can be applied to high-resolution orthophotos and which models are suitable for specific geospatial applications.

The main focus areas include:

  • The suitability of different foundation models for specific tasks
  • Transferability to high-resolution geospatial data
  • Fine-tuning with limited reference data
  • Generalizability across different regions

This creates a sound basis for deciding which AI approach is appropriate for a specific task.

Applications for Urban Development and Infrastructure

Geospatial foundation models are particularly interesting where large amounts of spatial image data need to be analyzed automatically and only limited training data is available.

Potential fields of application include:

  • Urban and regional development: Analyzing spatial structures and changes in high-resolution aerial imagery
  • Infrastructure-related analyses: Automatically detecting and evaluating relevant areas and structures
  • Land and potential analysis: Identifying suitable areas or specific spatial structures in geospatial data
  • Monitoring: Examining developments across different regions or time periods
  • Semantic segmentation: Automatically recognizing image regions and preparing them for further analysis

The Right AI for Your Geospatial Data

Foundation models are not automatically the best solution for every task. Key factors include the available data, its spatial resolution, existing reference data, and the desired analysis outcome.

Fraunhofer IGD is systematically building expertise to evaluate geospatial foundation models, select suitable models, and adapt them for specific use cases. This includes comparing different model approaches, fine-tuning with existing reference data, and technical integration for efficient use in your existing applications—for instance, through a reusable implementation foundation that allows suitable models to be integrated into existing geospatial applications more quickly.

You would like to know which AI approach is right for your geospatial data? 

Speak with our expert about your data and your specific use case.

Frequently Asked Questions about Geospatial Foundation Models

Expand all Close all

  • A geospatial foundation model is a pre-trained AI model for spatial data or remote sensing data. It can serve as a starting point for various analysis tasks and be adapted to specific applications.

  • Fraunhofer IGD specifically investigates high-resolution aerial imagery and digital orthophotos. In contrast, many existing geospatial foundation models were originally trained on lower-resolution satellite data.

  • Pre-trained models can reduce the need for application-specific training data. However, reference data and fine-tuning may still be required to adapt them to specific data and tasks.

  • Not fundamentally. Their suitability depends on the specific task and data foundation. Therefore, Fraunhofer IGD compares foundation models with established AI methods to determine which approach is best suited for each use case.