Denis060/sierraleone-agri-ml

First ML framework for crop yield prediction and post-harvest loss reduction in Sierra Leone — Evidence from FAOSTAT 2000–2024 # Can Machine Learning Forecast Rice Yields in Data-Constrained Settings? ## Satellite Climate Data, National Crop Statistics, and Lessons from Sierra Leone **Author:** Ibrahim Denis Fofanah **Affiliation:** Seidenberg School of Computer Science & Information Systems, Pace University, New York · RiseAfrica Foundation for STEM and Innovation, Sierra Leone **Email:** IF57774N@pace.edu --- ## Overview A reproducible pipeline that asks whether **rice yield (kg/ha)** in Sierra Leone can be forecast from publicly available data, under **strict anti-leakage discipline** and **walk-forward validation only**, benchmarked against naive baselines. The answer comes in two parts, both reported plainly: 1. **Crop statistics alone: no.** Trained on 25 years of FAOSTAT crop data (2000–2024), no ML model beats a simple persistence baseline (predict this year = last year). 2. **Adding free satellite climate data: yes.** With CHIRPS rainfall and NASA POWER temperature aggregated to national growing-season features, a climate-only XGBoost cuts forecast error by **one third** vs. persistence (RMSE 284 vs. 428 kg/ha) — a gain that holds for a linear model too and is robust to dropping the anomalous 2018 season. The dominant predictor is **May–June (planting-season) rainfall**, observable in CHIRPS months before harvest — the basis for a near-zero-cost early-warning capability for Sierra Leone's Ministry of Agriculture and Food Security. Honest boundaries, documented rather than hidden: **no model anticipated the 2018 yield collapse** (institutional, not climatic, in origin), and the record yields of 2020–2022 occurred in *below-average* rainfall years, consistent with input-driven policy gains. Full write-ups: outputs/model_report.md (v1) and outputs/model_report_v2.md (v2 + robustness). > An earlier version of this pipeline — same-year features, FAOSTAT aggregates, > random 70/30 split — produced an apparent R² of 0.96. Every component of that > number was leakage; it is preser …