EPS
← All batches·2605.08113

Do Foundation Model Embeddings Improve Cross-Country Crop Yield Generalisation? A Leave-One-Country-Out Evaluation in Sub-Saharan Africa

topic: othertop score: 100released: 2026-05-12first surfaced: 2026-05-12arXivPDFthreats2026-05-12

Authors: Yaw Osei Adjei

arXiv · PDF

Summary

The authors evaluate whether geospatial foundation model embeddings (Prithvi-EO and ViT-Base) improve crop yield prediction across countries in sub-Saharan Africa compared to traditional spectral features from Sentinel-2 satellite imagery. Using a leave-one-country-out cross-validation scheme on 6,404 maize fields across five countries, they find a large generalization gap: all methods achieve moderate R² within-country but universally negative R² cross-country. Frozen foundation model embeddings provide no meaningful advantage over hand-engineered spectral features, and the main bottleneck is distributional shift in yields across countries, not representation quality.

Main takeaways:

  • Within-country cross-validation overstates model performance; leave-one-country-out reveals near-total failure to generalize.
  • Geospatial foundation model embeddings (Prithvi-EO, ViT-Base) perform no better than traditional spectral features for cross-country yield prediction.
  • All feature sets achieve universally negative R² under cross-country testing, indicating the models are worse than a constant mean predictor.
  • The failure is attributed to yield distribution shift between countries, not poor feature quality.
  • The paper releases a reproducible negative benchmark to guide future work on cross-country agricultural prediction.

Relevance

Not related to my LLM work—this is about satellite-based crop yield prediction. Included because it's a foundation model paper, but the domain (geospatial vision) and problem (distribution shift in agriculture) have no overlap with persona installation, midtraining, or steering.

Threat model

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses limitation, negative, benchmark.

Abstract

arXiv:2605.08113v1 Announce Type: new Abstract: Accurate predictions of smallholder maize yields across national boundaries are critical for food security planning in sub-Saharan Africa, yet most published benchmarks report within-country performance that overstates true generalisability. This paper evaluates whether geospatial foundation model embeddings, specifically Prithvi-EO-1.0-100M and ViT-Base, outperform traditional Sentinel-2 spectral features under a Leave-One-Country-Out cross-validation scheme on 6,404 maize field observations from five African countries. The results show a clear generalisability gap: within-country random cross-validation yields moderate R^2 values, but all feature sets perform poorly under cross-country testing, with universally negative R^2. Frozen Prithvi-EO embeddings provide no meaningful advantage over engineered spectral features for cross-country prediction in this setting. The paper argues that the main limitation is a shift in yield distribution between countries rather than representation quality and releases a reproducible negative benchmark for future work.