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

Introduction: Sainfoin ( Methods: We conducted a multifactorial experiment to evaluate the depodding and dehulling efficiency of five commercially available sainfoin varieties. Seeds were processed using two different methods (belt thresher and impact dehuller) across five sample sizes. A pre-trained Faster R-CNN (Region-based Convolutional Neural Network) object detection model was fine-tuned to identify intact pods, whole seeds, and split seeds from images of the processed mixtures. These predictions were used to calculate processing efficiency (PE) for each variety. A comprehensive power analysis was performed to determine the minimum sample size of sainfoin pods required to detect differences in PE with high statistical power. Results: We observed strong varietal differences in PE, as well as clear effects of the processing method. Belt threshing produced mixtures with more intact pods, while the impact dehuller generated a higher proportion of split seeds. Increasing sample size led to more intact pods across all varieties and methods, and notably decreased seed proportion in belt-threshed samples. Statistical modeling combined with object detection outputs revealed that a minimum of 2 g of pods is required to reliably detect an absolute proportional difference of 0.25 in PE between two breeding lines with 80% power. Discussion: Our findings demonstrate that sainfoin varieties differ significantly in processing efficiency and that processing outcomes depend strongly on both method and sample size. Integrating deep learning-based phenotyping with robust statistical design enables efficient evaluation of processing traits and provides actionable guidelines for breeding programs. While deep learning models offer powerful, cost-effective tools for plant phenotyping, their outputs must be paired with rigorous statistical design to yield reliable and actionable insights for crop improvement.