FAIRgeo:
Fencing AI for Enhanced Reliability in Geo Services
AI on geo data is receiving high attention currently. However, literature that AI models are primarily reliable on their training data (e.g., 85%) and drastically unreliable outside it (e.g., 20%). The illustration shows a field fruit classification in the Netherlands, applying a model produced by Wageningen Research. It is clearly visible how the model hallucinates over water (bottom left), providing incorrect results.
In this case, a geographical restriction (e.g., based on cadastral data) would already be sufficient to constrain the model to valid situations. However, in general the challenge can be of arbitrary complexity and often not decidable at all.
Currently, (i) there is no general method for characterizing the validity of models in terms of spatial, temporal, and content-based criteria; (ii) there is no method for using a validity characterization to protect specific user requests; and (iii) there is no method to automate these mechanisms within a general spatiotemporal geodata infrastructure. Published models – for example on HuggingFace with currently 827,859 models – at best have a textual description of their applicability, often overly optimistic and generic, and currently not machine-readable.
We summarize the capability of a server to recognize invalid data/model combinations and react appropriately as "model fencing" to express that AI models during their inference get constrained to their individual application conditions ("comfort zone"). FAIRgeo attempts to structure the field and find decision criteria for selected situations.
Project Partners
Funding
Runtime 2025/03 - 2027/02, EU EFRE funding support 249922€.
|
