Abstract
The objective of the current study was to develop an original methodology based on deep learning algorithms and the Beer-Lambert law to estimate the average ear size and ear density at the microplot level for high-throughput field phenotyping. This methodology relies on RGB images viewing at nadir and at 45° taken at 1.5 m over the canopy by the Phenomobile V2 ground-robot. First, the YOLOv5 ear detection algorithm is applied to nadir RGB images to estimate ear density. Second, an ear segmentation algorithm (Unet CNN with Resnet 18 encoder) is applied to RGB images viewing at nadir and at 45° to compute the ears gap fraction at different viewing angles. The Beer-Lambert law (BL) is then applied to compute the ear area index (EAI) from the observed ear gap fraction. The light extinction parameter in BL is computed by using a geometrical model that represents ears as cylinders. The average ear size (the area of the vertical cross-section of the ear, in cm2) is then derived as the ratio between EAI and ear density.The methodology has been applied over a panel of 10 commercial bread wheat genotypes grown in 12 environments (including study sites, water stress treatments, sowing dates and initial plant density). Both, ear density and average ear size estimations were validated against destructive measurements. The relative RMSE is about 9% (i.e. 47 ears/m2) for ear density estimations and about 18% (i.e 1.3 cm2) for the average ear size. Furthermore, the estimated average ear size was strongly correlated with the average grain biomass per ear (r2=0.80 across genotypes and environments) measured at harvest and, similarly, the estimated EAI was also strongly correlated with grain yield (r2=0.83). Linear correlations between ear size and ear grain biomass per genotype (r2 ranging between 0.80 and 0.95) exhibit differences attributed to ear compactness, and suggest that ear size estimated from Phenomobile can be a good proxy for some yield components.