# Predicting crop yields with little ground truth: A simple statistical model for in-season forecasting

**Authors:** Nemo Semret  
**arXiv:** [2106.08720](https://arxiv.org/abs/2106.08720)  
**Date:** 2021-06-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2106.08720v3/

## Key points

- Fully automated model for in-season crop yield prediction
- Designed to address the lack of sub-national "ground truth" information
- Relies primarily on satellite data
- Incorporates careful feature engineering and a simple regression model
- Applicable in various locations worldwide
- Applied to 10 different crop-country pairs, including five cereals in Ethiopia and Kenya
- Achieves Root Mean Square Errors (RMSEs) of 5%-10% for predictions made nine months into the year and 7%-14% for predictions made three months into the year
- Provides daily forecasts for the final yield of the current year
- Trained using approximately 4 million data points for each crop-country pair, including historical country-level annual yields, crop calendars, crop cover information, Normalized Difference Vegetation Index (NDVI), temperature records, rainfall measurements, and evapotranspiration data.
- Utilizes satellite data to predict in-season crop yields accurately
- Can be effectively deployed in regions with limited ground truth information
- Impressive performance across multiple crops and countries demonstrates its potential as a valuable tool for agricultural forecasting and decision-making purposes.

## Summary

The study presents a fully automated model for in-season crop yield prediction, specifically designed to address the lack of sub-national "ground truth" information. The approach relies primarily on satellite data and incorporates careful feature engineering along with a simple regression model. This enables the model to be applicable in various locations worldwide. To evaluate the effectiveness of the model, it is applied to 10 different crop-country pairs, including five cereals (corn, wheat, sorghum, barley, and millet) in two countries (Ethiopia and Kenya). The results demonstrate that the model achieves Root Mean Square Errors (RMSEs) of 5%-10% for predictions made nine months into the year and 7%-14% for predictions made three months into the year. The model provides daily forecasts for the final yield of the current year. It is trained using a substantial amount of data, approximately 4 million data points for each crop-country pair. These data points consist of historical country-level annual yields, crop calendars, crop cover information, Normalized Difference Vegetation Index (NDVI), temperature records, rainfall measurements and evapotranspiration data. In summary, this research introduces an advanced automated model that utilizes satellite data to predict in-season crop yields accurately. By leveraging carefully engineered features and a straightforward regression approach, this model can be effectively deployed in regions where detailed ground truth information is limited. The impressive performance achieved across multiple crops and countries demonstrates its potential as a valuable tool for agricultural forecasting and decision-making purposes.