Abstract
In the early stages of selection, many plant breeding programmes still rely on visual evaluations of traits by experienced breeders. While this approach has proven to be effective, it requires considerable time, labour and expertise. Moreover, its subjective nature makes it difficult to reproduce and compare evaluations. The field of automated high-throughput phenotyping aims to resolve these issues. A widely adopted strategy uses drone images processed by machine learning algorithms to characterise phenotypes. This approach was used in the present study to assess the dry matter yield of tall fescue and its accuracy was compared to that of the breeder's evaluations, using field measurements as ground truth. RGB images of tall fescue individuals were processed by two types of predictive models: a random forest and convolutional neural network. In addition to computing dry matter yield, the two methods were applied to identify the top 10% highest-yielding plants and predict the breeder's score. The convolutional neural network outperformed the random forest method and exceeded the predictive power of the breeder's eye. It predicted dry matter yield with an R² of 0.62, which surpassed the accuracy of the breeder's score by 8 percentage points. Additionally, the algorithm demonstrated strong performance in identifying top-performing plants and estimating the breeder's score, achieving balanced accuracies of 0.81 and 0.74, respectively. These findings indicate that the tested automated phenotyping approach could not only offer improvements in cost, time efficiency and objectivity, but also enhance selection accuracy. As a result, this technique has the potential to increase overall breeding efficiency, accelerate genetic progress, and shorten the time to market. To conclude, phenotyping by means of RGB-based machine learning models provides a reliable alternative or addition to the visual evaluation of selection candidates in a tall fescue breeding programme.