cnongera/kenya-smallholder-yield-prediction

Kenya Smallholder Maize Yield Prediction # Kenya Smallholder Maize Yield Prediction **Author:** Clement Ongera Nyangoya **Date:** April 2026 **Location:** Nakuru County, Kenya (0.3066°S, 35.9668°E) ## Overview Pilot machine learning study predicting smallholder maize yield from CHIRPS satellite rainfall estimates. Conducted on the author's own farm across three growing seasons (2023–2024). ## Key Findings - **Dataset:** 3 plot-seasons from one smallholder farm in Kenya's Rift Valley - **Model:** Random Forest regression - **Top predictor:** Growing-season rainfall (`rain_growing_mm`, importance = 0.166) - **Performance:** R² = 0.774, RMSE = 296 kg/ha (training evaluation) - **Insight:** Rainfall during the growing season is the dominant yield-limiting factor, confirming farmer knowledge with quantitative evidence ## Data - **Yield records:** Farmer-reported maize yields (bag counts), planting dates, harvest dates from author's farm - **Rainfall:** CHIRPS daily precipitation estimates extracted via Climate Engine (0.05° resolution, ~5.5 km) - **Study site:** Semi-arid highland, ~1,900 m elevation ## Methods - Agronomic feature engineering: pre-season, growing-season, and flowering-period rainfall; dry-day count; peak daily rainfall; rainfall coefficient of variation - Random Forest regression with conservative hyperparameters (max_depth=3) to prevent overfitting on small sample - Feature importance analysis and predicted-vs-observed visualization ## Limitations - Small sample size (N=3 plot-seasons) — pilot study only - Farmer-reported yields (not formally weighed) - Single plot — no spatial generalization - CHIRPS resolution (5.5 km) much coarser than plot size - Training-on-all evaluation — no independent test set