Machine-Learning Crop Yield Forecasting for Smallholder Farms in the Nile Basin
Dr. Samuel Kefale · Department of Data Science & AIORCID: 0000-0002-1825-0097
Published 18 August 2026·Citation style: apa·521 views · 262 downloads
Abstract
Smallholder farms produce most of the food consumed across East Africa, yet yield forecasting tools designed for them are almost non-existent. We present a gradient-boosted ensemble that forecasts maize and teff yields up to 60 days before harvest using freely available satellite indices and minimal smartphone weather logs.
Introduction
Food security planning across the Nile basin depends on early, reliable yield forecasts for smallholder farms. Existing crop models presume data-rich commercial operations and agronomic records that smallholders rarely maintain. This study asks whether freely available satellite vegetation indices, combined with minimal crowdsourced weather observations from feature phones, can deliver forecast accuracy useful for cooperative price negotiation and regional planning.
Literature Review
Prior work in yield forecasting falls into three camps: process-based agronomic models, remote-sensing regressions and, more recently, deep-learning hybrids. Process models (e.g., DSSAT family) achieve strong accuracy but require field-level agronomic inputs. Remote-sensing approaches using NDVI and EVI indices have demonstrated skill at district scale, and gradient-boosted trees have emerged as robust performers on tabular agronomic data. Gaps remain in validating these approaches for teff, in integrating farmer-generated weather logs, and in testing whether forecasts lead to improved household outcomes rather than accuracy alone.
Methodology
We assembled three growing seasons (2022–2025) of plot-level yield observations for 1,240 smallholder farms in Oromia, combining Sentinel-2 derived vegetation indices, CHIRPS rainfall estimates, and daily temperature logs crowdsourced from 240 farmer smartphones. A gradient-boosted ensemble (XGBoost) with Leave-One-Village-Out cross-validation was benchmarked against ridge regression and a random forest. Feature importance was computed with SHAP values, and a farmer-facing SMS prototype was deployed with two cooperatives for live evaluation.
Results
The ensemble achieved 91% accuracy (R² = 0.83, MAE = 0.42 t/ha) in back-testing, outperforming the random forest by 6 percentage points and ridge regression by 21. SHAP analysis showed pre-flowering vegetation indices and cumulative rainfall to mid-June dominated predictions. In live deployment, 78% of surveyed cooperative members used the SMS forecast in price negotiation, and estimated household income uplift was 8–12% for maize sellers.
Discussion
The results demonstrate that accurate, actionable yield forecasting for smallholder farms does not require agronomic record-keeping or specialist equipment. The farmer-generated weather component materially improved accuracy in data-sparse areas, suggesting a scalable role for simple crowdsourcing. Limitations include site-specific calibration needs and the risk of over-reliance on single-model forecasts during anomalous seasons, which we mitigate with ensemble variance warnings.
Conclusion
Freely available satellite data plus minimal smartphone weather logs can forecast smallholder yields accurately enough to change market outcomes. We release our dataset and models openly and invite replication across other basins.
Acknowledgments
We thank the Oromia cooperative unions and 240 farmer enumerators, and the EIU Nile Basin Research Fund for financing this work.
Funding & Conflicts
Funded by the EIU Climate Adaptation in the Nile Basin seed grant (EA-2026-03). The funder had no role in study design or publication decisions.
Data Availability
Dataset and code are archived in the EIU digital repository under CC-BY-4.0.
References
- [1] Kefale, S. (2026). Machine-Learning Crop Yield Forecasting for Smallholder Farms in the Nile Basin. EIU Journal of Computing & Development, 4(2), 11–29.
- [2] Abera, Y., & Kefale, S. (2025). Crowdsourced weather for data-sparse agriculture. African Data Science Review, 8(1), 44–58.
- [3] Lobell, D. (2024). The use of satellite data for crop monitoring in food-insecure regions. Global Food Security, 33, 100–112.