Can Satellites Predict Yield? Ensemble Machine Learning and Statistical Analysis of Sentinel-2 Imagery for Processing...

Area of Science:

• Agricultural Science
• Remote Sensing
• Data Science

Background:

• Accurate processing tomato yield prediction is crucial for agricultural management.
• Satellite imagery offers a scalable solution for crop monitoring.
• Vegetation indices (VIs) derived from satellite data can indicate crop health and predict yield.

Purpose of the Study:

• To develop a robust method for predicting processing tomato yield.
• To evaluate the effectiveness of Sentinel-2 derived VIs at different phenological stages.
• To compare statistical analysis with AutoML techniques for yield prediction.

Main Methods:

• Utilized Sentinel-2 satellite imagery to calculate five VIs at 5-day intervals during the 2021 growing season.
• Collected ground-truth yield data from 108 processing tomato fields in central Greece.
• Applied statistical analysis (Pearson correlation) and open-source AutoML (ARD regression, SVR) for yield prediction.

Main Results:

• Highest correlations between VIs and yield were observed between 80 and 90 days of the growing season (r = 0.72–0.75).
• AutoML models confirmed peak performance during this period, with adjusted R² values ranging from 0.60 to 0.72.
• An ensemble model combining ARD regression and SVR achieved the highest precision (adj. R² = 0.67 ± 0.02).

Conclusions:

• Sentinel-2 VIs are effective predictors of processing tomato yield, particularly during mid-to-late growth stages.
• AutoML techniques, especially ensemble methods, enhance prediction accuracy.
• This approach offers a robust and scalable tool for agricultural yield forecasting.