Assessment of stay‐green in grain sorghum across multi‐location trials using unoccupied aerial systems and machine learning

Abstract

Sorghum ( Sorghum bicolor (L.) Moench) is a diploid crop that is primarily grown in regions with high temperatures and limited irrigation. Breeding efforts often focus on stay‐green, a trait associated with green leaf area after the crop experiences post‐flowering drought. When present, this trait is assessed using subjective visual categorical ratings. The objective of this study was to assess stay‐green across nine environments and to develop a machine learning model that can estimate stay‐green using vegetative indices collected by an unoccupied aerial system. Hybrids were derived from a partial factorial design that included material from three public and one private breeding program. Results showed that each program had heritable variation for this trait and the benefit of sharing material to develop top‐performing stay‐green hybrids varied among each breeding program. Five machine learning models were built to estimate stay‐green; they produced accuracies up to 92% (gradient boosting model) in testing and training. The validation dataset, which contained data from four unseen environments, produced average accuracies up to 66% (naïve Bayes model). The naïve Bayes model stay‐green estimates also resulted in a heritability 16% higher than that of the field‐collected ratings. Overall, the naïve Bayes and gradient boosting models performed best, demonstrating potential for more efficient and less biased stay‐green rating collection in grain sorghum.