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

Short summary

A Leave-One-Country-Out evaluation of 6,404 maize field observations across five sub-Saharan African countries finds that foundation model embeddings (Prithvi-EO-1.0-100M, ViT-Base) provide no meaningful advantage over traditional Sentinel-2 spectral features for crop yield prediction—all approaches achieve negative R² under cross-country testing. The authors attribute poor generalization to distribution shifts in yield data between countries rather than representation quality, and release a reproducible negative benchmark to guide future agricultural AI research.

• •Foundation model embeddings fail to outperform hand-engineered features on cross-country crop yield prediction
• •Both feature sets achieve negative R² under Leave-One-Country-Out evaluation across 6,404 maize fields in 5 African countries
• •Distribution shift between countries is the limiting factor, not representation quality—authors release benchmark for future work

Generated with AI, which can make mistakes.