Early cereal yield prediction using machine learning and Sentinel-1 & Sentinel-2 satellite data

Abstract

Forecasting cereal production is crucial for food security, especially in southern Mediterranean countries like Tunisia that rely on cereal imports. While optical indices are commonly used for cereal yield forecasting, integrating radar data within Machine learning (ML) models for simultaneous grain and straw yield prediction remains significantly underexplored. This study develops early prediction models for cereal yields (grain and straw) at the field scale across the Kairouan plain, central Tunisia. For this purpose, multi-sources data are used, including remote sensing indices from Sentinel-1 and Sentinel-2. Cereal test plots were monitored during 2017–2018 and 2020–2021, and field measurements, such as grain and straw yields, were also collected. For model development and validation, we split the data 70:30 using Linear Regression (LR) models and ML models. Other models were tested, but the Random Forest (RF) model outperformed the rest in estimating early grain yields, with an R2 of 0.91 and a nRMSE of 10.2%. Conversely, LR model showed strong correlations for straw yield (R2 = 0.76, nRMSE = 18.2%). Integrating multiple data sources including essential satellite indices not only improves prediction accuracy over single-source models but also enables early grain and straw yield forecasts, supporting decision-making more than three months before harvest.