Prithvi-EO fails cross-country crop yield prediction in Sub-Saharan Africa study

A new study from researcher Yaw Osei Adjei puts foundation models for agriculture to a rigorous test. The paper, "Do Foundation Model Embeddings Improve Cross-Country Crop Yield Generalisation?" evaluates whether NASA's Prithvi-EO-1.0-100M and the standard ViT-Base, both large vision models trained on satellite imagery, can generalize maize yield predictions across five Sub-Saharan African countries. Using a Leave-One-Country-Out cross-validation on 6,404 field observations, the study simulates real-world conditions where a model trained on data from some countries must predict yields in an unseen country. The benchmark compares frozen embeddings from these foundation models against traditional Sentinel-2 spectral features like NDVI.

The results are stark: all models fail at cross-country prediction, yielding universally negative R^2 values—meaning they perform worse than simply predicting the mean yield. Within-country validation showed moderate R^2, but that performance evaporated under cross-country testing. Importantly, frozen Prithvi-EO embeddings provided no improvement over simpler spectral features. The author identifies the core issue as a shift in yield distributions between countries, not representation quality of the embeddings. This suggests that foundation models alone cannot overcome domain shift in agricultural applications. The paper releases code and processed results as a reproducible negative benchmark, challenging the assumption that large geospatial models automatically improve generalization. For practitioners, this highlights the need for methods that address distribution shift directly, such as domain adaptation or transfer learning, rather than relying on better representations.