Computer vision associated to multivariate genomic selection methods increase yield prediction accuracy in blueberry

Blueberry (Vaccinium spp.) is among the most consumed soft fruits and an important source of health-promoting compounds. Among the key traits driving selection in breeding programs, yield is the most important. The standard way to measure yield is harvesting and weighing the total number of berries, a process that is laborious, expensive, prone to measurement errors, and not scalable to short production windows. To circumvent this, breeders rely on visual scores, an approach that offers scalability but includes subjectivity. In this study, we investigated the use of computer vision methods for fruit detection to guide breeding decisions. Our fundamental hypothesis is that integrating machine learning and molecular breeding could strengthen genetic analyses and support molecular breeding. To test it, a large blueberry breeding population was evaluated using different yield-related metrics, including fruit detection via computer vision, visual scores in different phenological stages, and total berry weight. Our contributions are threefold: (i) using computer vision, we better assessed yield potential, producing genetic parameters that improved residual control and leveraged genetic variation; (ii) we inferred the genetic basis of yield in blueberry and highlighted the importance of non-additive effects on phenotypic expression; and (iii) we showed that computer vision and visual scores combined in multivariate genomic prediction models resulted in better predictive abilities. Altogether, for the first time in the blueberry literature, we demonstrated how computer vision and molecular breeding can be integrated in the same framework to guide breeding decisions.