Leveraging High-dimensional Seed Colorspace in Phenomic Selection Models for Herbaceous Perennial Crops

Abstract

Herbaceous perennial crops, which only recently entered the domestication pipeline, offer the potential of ecological benefits in agricultural systems but remain underdeveloped, partly due to slow breeding cycles. Early-stage phenomic selection, the process of using high-dimensional secondary traits to predict target traits, may accelerate improvement. We evaluated whether seed traits derived from image scans could be used to predict germination traits in three perennial crop candidate species: sainfoin (Onobrychis viciifolia), intermediate wheatgrass (Thinopyrum intermedium), and silflower (Silphium integrifolium). Over 20,000 seeds were scanned, and PlantCV was used to extract 692 hue, saturation, and value (HSV) color features along with five morphological traits. We demonstrated that seed color, morphology, and germination traits were all influenced by maternal family, but patterns of correlation between germination traits and color and morphology traits are complex. To enable prediction from high-dimensional seed traits, we constructed relationship matrices from combinations of HSV and morphology features for use in phenomic selection models to predict germination proportion and timing. Model performance varied by trait, species, and cohort, and in many cases, predictions were significantly better than chance. HSV features, when combined with morphology, often yielded the most accurate predictions, with performance reaching a maximum of r = 0.33. These results demonstrate the potential of low-cost, image-based phenomic data to inform early-stage selection and support the development of sustainable perennial crops. Plain language summaryMany of the worlds most important crops (e.g., wheat, maize, soy) live for less than one year, leaving soil bare for many months and subject to erosion. Longer-lived (perennial) alternatives to these crops have the potential to support a more sustainable agricultural system by improving soil health and reducing the need for chemical inputs. However, perennial herbaceous plants were not domesticated by early farmers, and now lag behind annual crops in development. They take longer to breed and often lack the genetic resources needed for modern breeding approaches. In this study, we tested whether simple seed images derived from color scanners could be used to predict key traits like germination proportion and timing in emerging perennial crops. We found that seed color and shape varied by maternal family and that color, especially when combined with seed size and shape, could help predict germination performance. These predictions were significantly better than chance. This work shows that low-cost, image-based seed data can support early-stage selection in breeding programs, especially for undeveloped perennial crops, helping save time, reduce cost, and accelerate progress toward more sustainable agriculture.