Harvest Science
Computer Vision for Harvest Timing: Predicting Grain Moisture From Canopy Imagery
L. Mbeki, S. Haruna, P. Rossi · Farm Better Research Lab
January 2026 · 11 min read
31.4k
Images labelled
±1.4 pp
Moisture error
91%
Agronomist agreement
Abstract
A convolutional model estimates maize grain moisture from smartphone canopy photographs, giving farmers a harvest window without laboratory sampling.
Method
31,400 labelled canopy images were paired with destructive moisture tests across three agro-ecological zones. Images were captured on entry-level smartphones under uncontrolled lighting to reflect real field conditions.
Findings
- Mean absolute error of 1.4 percentage points against laboratory grain moisture.
- Recommended harvest windows aligned with agronomist judgement in 91% of cases.
- Early-harvest losses reduced by an estimated 7% of yield in the trial cohort.
- Performance dropped sharply below 200 lux, suggesting a simple in-app lighting check.
Conclusion
Smartphone imagery is a credible substitute for moisture meters for harvest scheduling, provided the app enforces basic capture quality.
Note: this article is an illustrative mock-up created for the Farm Better site and is not a peer-reviewed publication.