Using machine learning for image-based analysis of sweetpotato root sensory attributes.

Area of Science:

• Agricultural Science
• Computer Science
• Food Science

Background:

• Sweetpotato breeding relies on evaluating sensory traits like color and mealiness for consumer acceptance.
• Current sensory evaluation methods using human panels are slow, costly, and limit breeding throughput.
• High-throughput methods are needed to accelerate the selection of desirable sweetpotato varieties.

Purpose of the Study:

• To develop and validate machine learning models for predicting sweetpotato flesh color and mealiness using image analysis.
• To create a high-throughput phenotyping tool to assist sweetpotato breeders.
• To augment traditional sensory evaluation methods with objective, automated analysis.

Main Methods:

• Captured images of boiled sweetpotato cross-sections using the DigiEye imaging system.
• Applied image pre-processing for background elimination and feature extraction.
• Developed and trained machine learning models (Linear Regression, Random Forest, Gradient Boosting) to predict sensory attributes against human panel data.

Main Results:

• Machine learning models achieved high accuracy for predicting flesh color (R² up to 0.92) and mealiness (R² up to 0.85).
• Model performance met the desired R² threshold of 0.80, indicating comparability to human sensory panels.
• The developed system was successfully deployed and tested by breeders at the International Potato Center.

Conclusions:

• Image-based machine learning offers a high-throughput, accurate solution for evaluating sweetpotato flesh color and mealiness.
• This approach can significantly accelerate the sweetpotato breeding cycle by enabling faster selection of promising varieties.
• Automated sensory attribute prediction has the potential to increase the adoption rate of new sweetpotato varieties by consumers.