Predicting Biomass and Yield in a Tomato Phenotyping Experiment Using UAV Imagery and Random Forest

Area of Science:

• Agricultural Science
• Plant Science
• Remote Sensing

Background:

• Biomass and yield are critical for agricultural system assessment.
• Predicting these variables at the farm scale is challenging, especially with abiotic stresses like salinity.
• Wild tomato (Solanum pimpinellifolium) serves as a model for evaluating phenotyping strategies.

Purpose of the Study:

• To predict fresh shoot mass, fruit number, and yield mass in tomato plants using UAV-derived data.
• To assess prediction accuracy for control versus salt-stressed plants.
• To determine the optimal timing for UAV data collection for yield prediction.

Main Methods:

• Field and UAV-based phenotyping of 1200 Solanum pimpinellifolium plants (600 control, 600 salt-treated).
• Collection of RGB and multispectral UAV imagery at various intervals before harvest.
• Application of a random forest machine learning model using UAV-derived features (shape, vegetation indices, texture).

Main Results:

• Shape features (plant area, border length, width, length) were most important for prediction.
• Multispectral imagery 2 weeks pre-harvest yielded high explained variances: 87.95% for shoot mass, 63.88% for fruit number, 66.51% for yield mass.
• Prediction accuracy for salt-stressed plants improved when control plants were excluded; control plant prediction was unaffected by salt-stressed plant inclusion.
• Average biomass and yield predicted within 4.23% of field measurements up to 8 weeks pre-harvest.

Conclusions:

• UAV-based phenotyping with machine learning enables accurate yield prediction in tomato.
• This method is effective for both healthy and salt-stressed plants, with distinct model performance observed between groups.
• Findings support improved yield forecasting for agricultural planning and management.