SciGroveBeta
Climate

How Usable Are Geospatial Foundation Models? A Systematic Evaluation of 89 Models

Robin Young, Artyom Gabtraupov, Kenzy Soror, Srinivasan Keshav

Featured August 9, 2026

Original Paper

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

Powerful AI models for Earth observation are hard for scientists to use because they lack simple interfaces and don't show how sure their predictions are, making it tough to trust them for important decisions.

In depth
The paper addresses the critical gap between the technical sophistication of Geospatial Foundation Models (GeoFMs) and their practical usability by domain experts. The authors developed a seven-dimension evaluation framework, grounded in Human-Computer Interaction (HCI) theory and expert input, to systematically assess 89 GeoFMs. Their findings reveal significant accessibility gaps and a profound misalignment in trust infrastructure, where crucial features like uncertainty quantification are almost entirely absent.

Key Takeaways

  • 1
    Over half of the surveyed GeoFMs are largely inaccessible to practitioners, requiring deep machine learning expertise or pre-training from scratch, indicating a severe accessibility cliff.
  • 2
    There is a critical trust gap: while nearly all models provide replicable benchmarks, zero models offer documented uncertainty quantification, which domain experts prioritize for trusting outputs.
  • 3
    The paper introduces a novel seven-dimension evaluation framework (covering Access, Interaction, Trust, Community, Permanence, Multilingual, Offline) that serves as both a discriminator for individual models and a diagnostic for systemic field-level gaps.

Conceptual Flow

HIGH LEVEL
1
Methodology: How GeoFMs Were Evaluated

The study asked scientists what they needed, then created a checklist to see how easy 89 Earth-watching AI models were to use.

Talk to Scientists
Build Checklist
Rate AI Models
2
Results: Key Usability Gaps Found

Most Earth-watching AI models are hard to start using, and almost none show how confident their predictions are, which scientists really need.

Hard to Start Using
Plus
No Confidence Info