AI and Datacubes - a Happy Marriage?

Datacubes are an accepted cornerstone for Analysis-Ready Data (ARD). One analysis technique of skyrocketing importance today is AI, and this begs the question: how to generalize evaluation of pre-trained models on datacubes?

From a theoretial viewpoint, the connection is immediate: datacubes mathematically resemble tensors, and EO models evaluate tensors. In practice, though, the situation is less straightforward as our experiments with different models have shown. A main issue is the variety and the lack of standardized interfaces of ML models: different input data are processed, data need model-specific preprocessing, and several more. In our research towards offering ML-on-datacubes as a commodity in a federated datacube infrastructure we have collected challenges and methods for presentation.

In our demo, we present AI-Cubes as an enabling concept uniting AI and datacubes. The demos will approach the theme from two sides:

- AI support for datacube query writing: We have trained a chatbot to explain and assist with datacube queries in the OGC/ISO/INSPIRE WCPS standard. This can act as a productiity-enhancing tool for both expert and non-expert users. We demonstrate live how specific questions get answered, such as phrasing NDVI on Sentinel-2 data.

- AI model evaluation on datacubes: particularly attractive is that datacubes allow simple navigation to any area, any time, and even from federated services. This we demonstrate live.

We also highlight challenges coming with this simple data access: models do not convey the same performance anywhere, anytime. This has led to new research on "model fencing", ie: attempting to restrict model application to situations where they exhibit sufficient accuracy. We present first ideas of this research.

Altogether, we cast light on the combination of datacubes and AI from a service and infrastructure perspective.