Adedeji-Taiwo/maize-yield-forecast

End‑to‑end reproducible pipeline for subnational maize yield forecasting in Nigeria using Earth observation and machine learning (Random Forest, LOYO, SHAP). # Maize Yield Forecasting Pipeline for Nigeria ### Earth Observation + Machine Learning · Fully Reproducible This repository contains the complete, end-to-end computational pipeline described in the paper: > **Operational Value of Earth Observation and Reanalysis Predictors for Subnational Maize Yield Forecasting in Nigeria** > *Taiwo Adedeji Michael, Adeyeye Daniel Oludayo, Oladeji Qudus Olamide, and Olaniyi Bolaji Samuel* > Preprint available at Operational Value of Earth… --- ## What this repo does This pipeline builds a **19-year state-level panel** (2004–2022) for Nigeria by merging: - **HarvestStat Africa** – subnational maize yield statistics - **MODIS NDVI (MOD13A3)** – cropland-masked phenology - **CHIRPS** – dekadal rainfall - **ERA5-Land** – growing-season air temperature - **MODIS LST (MOD11A2)** – daytime land surface temperature - **SoilGrids 2.0** – static soil properties A **Random Forest** model is evaluated under strict **Leave-One-Year-Out (LOYO)** cross-validation with a **predictor-block ablation** design. The ablation quantifies the marginal predictive value of each data stream, with a particular focus on whether MODIS LST adds any operational skill beyond ERA5-Land air temperature. **Key findings:** - MODIS LST added **negligible** forecast skill (ΔRMSE = 0.001 t ha⁻¹), explained by high collinearity with ERA5-Land (r = 0.89). - A simple persistence baseline (prior-year yield) **outperformed** the full EO-only model. - The best EO-only model (NDVI + CHIRPS + ERA5) achieved RMSE = 0.475 t ha⁻¹, MAPE = 29.4%. - Adding soil properties and a lagged yield term reduced RMSE to 0.407 t ha⁻¹, MAPE = 22.6%. - Forecast skill varied sharply across agro-ecological zones and was structurally bounded by yield variability. --- ## Repository structure ```bash . ├── data/ # input CSVs (yield, NDVI, rainfall, temperature, LST, soil) ├── results/ # all generated tables and CSVs ├── figures/ …