Recently, remote sensing large models and geospatial foundation models have attracted widespread attention. Apart from the commercial models of ArcGIS and Google I have introduced before, there are numerous open-source alternatives such as SatMAE, Scale-MAE, Prithvi, Clay, Tessera, DOFA and more. How capable are these models, and can they be deployed in practice? A research team from the University of Cambridge has published a usability evaluation paper, and I will walk through the findings in this article.

Paper Background

Introductions to geospatial foundation models (GeoFMs) are usually filled with technical jargon: network architecture, pre-training datasets, parameter counts and benchmark scores. Such information is easy to interpret for machine learning practitioners, yet business users with limited relevant experience face substantial barriers to running these models locally.

The team conducted preliminary expert interviews and defined seven evaluation dimensions. Two raters systematically assessed 89 geospatial foundation models using a unified scoring rubric, covering models released between October 2025 and January 2026.

Evaluation Dimensions

The team established seven evaluation dimensions as follows:

  1. Accessibility
  2. Customization and Fine-Tuning
  3. Trustworthiness: reproducible benchmarks and interactive verification
  4. Community and Technical Support
  5. Scientific Longevity & Reproducibility: open-source availability, DOI archiving, and service stability
  6. Multilingual Support
  7. Offline Usability

Accessibility is the most fundamental and critical dimension, further divided into six tiers from Level 0 to Level 5:

  1. Level 0: Source code only, no pre-trained weights
  2. Level 1: Pre-trained weights available, yet complex environment setup required
  3. Level 2: Environment configuration documents including Dockerfile / requirements files
  4. Level 3: Annotated notebooks or complete demonstration scripts with sample workflows
  5. Level 4: Hosted API access
  6. Level 5: Graphical interfaces including web clients, desktop software, or QGIS plugins

Over Half of the Models Remain Hard to Adopt

We first look at accessibility distribution shown in the figure below. Among the 89 models, 26 fall into Level 0, offering only source code without weights and cannot be initialized. Another 19 models are rated Level 1, requiring complicated environment configuration. Combined, these two tiers account for 45 models, exceeding half of the total. Only six models reach Level 4 or Level 5: MOSAIKS, Google AlphaEarth, Prithvi (via the IBM/NASA TerraTorch ecosystem), AIEarth, SatLas, and Tessera.

Among the 63 models rated above Level 1, 61 can reproduce benchmark results published in their original papers. 52 provide documented inference pipelines for running data, while 41 support fine-tuning. Prompt-based exploration and human-in-the-loop retraining are rare, with only three models supporting each function. In terms of community support, roughly 44 models rely merely on README files; approximately eight have independent documentation portals, tutorials or active user communities. Tessera and Prithvi stand out as two projects with relatively complete documentation, tutorials and community ecosystems.

From a temporal perspective, model releases saw clear growth in 2023 and 2024. Nevertheless, the proportion of Level 0 models (without pre-trained weights) persists at around one quarter, and very few models provide GUI or API access.

Conclusion

Ordinary users have straightforward requirements: accessible data, model compatibility with familiar platforms including QGIS and Earth Engine, and verifiable analytical outputs. Currently, most models demand manual environment configuration, custom scripting and data preprocessing, not to mention comprehensive documentation and graphical user interfaces.

In my opinion, most geospatial foundation models are research outputs released to follow the foundation model trend. Apart from a small number of candidates such as AlphaEarth, Tessera and Prithvi, the rest are still far from direct deployment in real-world projects.

References

  1. Paper abstract page: https://arxiv.org/abs/2608.03804v1
  2. Full HTML version: https://arxiv.org/html/2608.03804v1
  3. Paper PDF: https://arxiv.org/pdf/2608.03804v1