Yield Modelling
Season-Ahead Yield Prediction by Fusing Satellite Indices With Farm Management Records
C. Nakamura, B. Kimathi, E. Duarte · Farm Better Research Lab
May 2025 · 12 min read
2,100
Fields modelled
11.6%
Forecast error
18.2%
Baseline error
Abstract
We test whether adding farmer-entered management records to satellite vegetation indices improves season-ahead yield forecasts for maize and soybean.
Method
Sentinel-derived NDVI and EVI time series for 2,100 fields were fused with planting dates, variety, and fertiliser records in a gradient-boosting model, benchmarked against a satellite-only model.
Findings
- Forecast error improved from 18.2% to 11.6% MAPE when management records were included.
- Planting date was the single most informative non-satellite feature.
- Forecasts stabilised roughly 55 days after emergence.
- Fields with incomplete records saw no improvement over the satellite-only baseline.
Conclusion
Data quality from farmers, not satellite resolution, is the binding constraint on yield forecasting accuracy.
Note: this article is an illustrative mock-up created for the Farm Better site and is not a peer-reviewed publication.