rasdaman News Archive

rasdaman newsletter 10/2025

FAIRgeo: Towards Safe and Simple Geo-AI

September 2025 -- While recent advances in AI/ML offer promising avenues for Earth data exploitation, an effective application is hindered by complexity and the risk of applying models to inappropriate data, leading to inaccurate or misleading results. The FAIRgeo project establishes methods allowing Big Earth Data platforms to offer third-party ML models to expert and non-expert users in a simple and safer way, based on the novel concept of model fencing. Now the first project review has confirmed the FAIRgeo approach and its first results.

Datacubes are acknowledged as a cornerstone for analysis-ready data as they allow more intuitive, human-centric services - they abstract away technical pains. However, new technical pains arise with AI:

Difficult to use, complex preprocessing, insufficiently documented -- ML model models require Python wrappers which often are complex and need significant knowledge about data encodings and the processing environment. Experts and non-experts alike should be able to use ML conveniently.
While datacubes make it trivially simple to access any region and any time in the cube, models do not perform satisfactorily everywhere. In fact, it has been observed that accuracy can degrade to 20% or less. Users should be protected against mis-application.

These novel features combine seamlessly with existing rasdaman functionality, such as location-transparent federation, fine-grain security, intelligent automated data import, and more. The project is fully committed to the interoperable Coverage standards of ISO, OGC and EU INSPIRE, in particular the Coverage Implementation Schema (CIS) data model, the Web Coverage Service (WCS) service model with its Web Coverage Processing Service (WCPS) geo datacube query language.

Interestingly, first experiments with AI-Cubes in the engine have shown a very satisfactory speed, significantly outperforming even standard Python workflows.