Comparison of Mask-R-CNN and Thresholding-Based Segmentation for High-Throughput Phenotyping of Walnut Kernel Color

Area of Science:

• Plant Science
• Computer Vision
• Agricultural Technology

Background:

• High-throughput phenotyping is crucial for modern plant breeding, improving efficiency over subjective traditional methods.
• Machine learning (ML) computer vision, particularly convolutional neural networks (CNNs), offers advanced image segmentation for plant trait analysis.
• Quantitative image analysis provides objective data, essential for accurate plant breeding selection.

Purpose of the Study:

• To compare the performance of rule-based thresholding and a Mask-R-CNN instance segmentation pipeline for walnut kernel phenotyping.
• To evaluate the correlation between quantitative image analysis methods and human evaluations of walnut kernels.
• To assess the consistency and adaptability of different image analysis techniques in plant breeding.

Main Methods:

• Collected over 90,000 walnut kernels from 3000 trees over three years.
• Employed rule-based thresholding (magick package in R) and a Mask-R-CNN pipeline for image analysis.
• Compared quantitative data (lightness, size) from image analysis with two sets of human evaluations.

Main Results:

• Both image analysis methods showed high correlation for lightness (r² = 0.997) and size (r² = 0.984).
• The CNN method required minimal training (13 images) and was robust to image staging variations, unlike thresholding.
• Human evaluations showed low correlation with image analysis and with each other, highlighting consistency issues.

Conclusions:

• Pixel classification offers consistent, quantitative data for plant phenotyping, surpassing subjective human assessments.
• The Mask-R-CNN approach is adaptable for imperfectly staged images and can be retrained for complex traits like spotting and shrivel.
• Quantitative image analysis, especially ML-based methods, significantly enhances the objectivity and efficiency of plant breeding programs.