Deep learning driven, image-based phenotyping of seed processing efficiency in sainfoin (Onobrychis viciifolia)

Area of Science:

• Agricultural Science
• Plant Breeding
• Computational Biology

Background:

• Sainfoin (Onobrychis spp.) is a traditional forage legume with emerging potential as a perennial grain crop.
• Limited research exists on breeding sainfoin for superior grain processing traits.

Purpose of the Study:

• To evaluate depodding and dehulling efficiency across five sainfoin varieties.
• To assess the impact of processing methods (belt thresher, impact dehuller) and sample sizes on processing efficiency.
• To develop a deep learning-based phenotyping approach for evaluating sainfoin processing traits.

Main Methods:

• A multifactorial experiment tested five sainfoin varieties with two processing methods and five sample sizes.
• A Faster R-CNN object detection model was fine-tuned to quantify processing outcomes.
• Power analysis determined the minimum sample size for reliable efficiency difference detection.

Main Results:

• Significant varietal differences in processing efficiency (PE) were observed.
• Belt threshing yielded more intact pods; impact dehulling produced more split seeds.
• A minimum of 2g of pods is needed to detect a 0.25 PE difference with 80% power.

Conclusions:

• Sainfoin variety, processing method, and sample size critically affect grain processing outcomes.
• Deep learning phenotyping combined with statistical design offers efficient evaluation for breeding programs.
• Rigorous statistical design is essential for reliable insights from AI-driven phenotyping in crop improvement.