yoadjei/yield-africa

# Yield Africa — Reproducibility Package **Paper:** *Do Foundation Model Embeddings Improve Cross-Country Crop Yield Generalisation? A Leave-One-Country-Out Evaluation in Sub-Saharan Africa* **Author:** Yaw Osei Adjei, Department of Computer Science, KNUST, Ghana --- ## Overview This repository contains the full code pipeline to reproduce all experiments and figures in the paper. The pipeline evaluates 18 experimental conditions (3 feature sets × 3 regressors × 2 CV schemes) for smallholder maize yield prediction across Kenya, Malawi, Nigeria, Rwanda, and Tanzania. **Key result:** All LOCO R² are negative. Frozen Prithvi-EO embeddings do not outperform 10-band Sentinel-2 spectral features for cross-country yield prediction. --- ## Repository Structure ``` yield_africa/ ├── scripts/ │ ├── 01_download.py # download GROW-Africa labels from Zenodo │ ├── 01b_gee_extract.py # export S2 patches via Google Earth Engine │ ├── 01c_chirps.py # extract CHIRPS rainfall features │ ├── 01d_harveststat.py # merge HarvestStat Africa (Nigeria coverage) │ ├── 01e_sample.py # stratified sampling for GEE export │ ├── 02_preprocess.py # build master_dataset.parquet │ ├── 03_extract_embeddings.py # extract Prithvi-EO and ViT-Base embeddings │ ├── 04_train_eval.py # train + evaluate all 18 conditions │ ├── 05_figures.py # generate all paper figures │ └── prithvi_mae.py # Prithvi-EO model architecture (from HF repo) ├── data/ │ ├── raw/ # raw downloads (not tracked by git) │ └── processed/ # results_all.csv, results_loco_country.csv (tracked) ├── figures/ # all 6 paper figures (PDF) ├── models/ # Prithvi model weights (not tracked — download below) ├── paper/ │ ├── main.tex # LaTeX source │ └── references.bib # BibTeX references ├── requirements.txt └── run_all.sh # end-to- …