Prediction of Early Vigor from Overhead Images of Carinata Plants

Abstract

Breeding more resilient, higher yielding crops is an essential component of ensuring ongoing food security. Early season vigor is signi cantly correlated with yields and is often used as an early indicator of tness in breeding programs. Early vigor can be a useful indicator of the health and strength of plants with bene ts such as improved light interception, reduced surface evaporation, and increased biological yield. However, vigor is challenging to measure analytically and is often rated using subjective visual scoring. This traditional method of breeder scoring becomes cumbersome as the size of breeding programs increase. In this study, we used hand-held cameras tted on gimbals to capture images which were then used as the source for automated vigor scoring. We have employed a novel image metric, the extent of plant growth from the row centerline, as an indicator of vigor. Along with this feature, additional features were used for training a random forest model and a support vector machine, both of which were able to predict expert vigor ratings with an 88:9% and 88% accuracies respectively, providing the potential for more reliable, higher throughput vigor estimates.