Weather Intelligence
Hyper-local Rainfall Forecasting for Smallholder Plots Using Ensemble Machine Learning
A. Wanjiru, D. Okoth, M. Ferreira · Farm Better Research Lab
March 2026 · 9 min read
480
Plots studied
88%
Forecast accuracy
61%
Baseline accuracy
Abstract
We evaluate an ensemble of gradient-boosted trees and a lightweight temporal transformer trained on satellite reanalysis, ground station data and low-cost farm sensors to forecast rainfall at 1 km resolution across 480 smallholder plots.
Method
Four seasons of data were collected from 480 plots between 1 and 6 hectares. Model inputs combined 10-day satellite reanalysis windows, hourly station readings and on-farm humidity and barometric sensors. Baseline was the national 12 km regional forecast.
Findings
- 72-hour rain onset accuracy improved from 61% (regional baseline) to 88% at plot level.
- False-alarm rate for spray-window advisories dropped by 34%.
- Farmers who acted on the advisories reported 19% fewer wash-off pesticide re-applications.
- Model degrades gracefully: with sensors removed, accuracy falls to 79%, still above baseline.
Conclusion
Plot-level forecasting is achievable without dense instrumentation. The largest gains come from fusing cheap on-farm sensors with existing public satellite products rather than from model complexity.
Note: this article is an illustrative mock-up created for the Farm Better site and is not a peer-reviewed publication.