Résumé
Digital phenotyping has the potential to sustain the use of genetic resources for fruit tree breeding by increasing the throughput and the accuracy of phenotypic measurements. The aim of this study is to assess the potential of RGB cameras to measure trunk circumference and flower density in apricot and peach. Two vectors were employed to this aim: the ground-based phenotyping pole LITERAL equipped with two RGB cameras, and an unmanned aerial vehicle (UAV) AUTEL Evo II equipped with one RGB camera. We screened a large diversity of peach and apricot accessions planted in a core-collection design within orchards containing 2 to 5 randomized blocs of 206 and 150 genotypes, respectively. For estimating trunk circumference, we used pictures from LITERAL combined with wood segmentation and stereovision algorithms. Although correlation between observed and predicted values for trunks between 10 and 40cm circumference were high (r2= 0.66), trunk estimation errors were comprised between 9.5 and 13.5mm which is insufficient to measure interannual growth. Further analyses revealed that this lack of precision essentially stems from wood segmentation errors which can be addressed by improving the training set. Flower density was estimated either with single flower detection (LITERAL images) or with photogrammetric analyses followed by flower and wood segmentation (UAV images). We discuss the precision and feasibility of these measurements and our strategy to produce analytic pipelines in order to deploy digital phenotyping for trunk circumference and flower density traits which are of major interest for breeding.