Résumé
Assessing flowering intensity in orchards is crucial for effective tree management, yield optimization and to forecast adaptation of cultivars in the context of climate change. Flowering intensity is influenced by genetic and environmental factors. For now, the traditional methods to assess flowering intensity are based on visual observation and are tedious and qualitative. Our aim was to develop an image analysis pipeline to estimate as accurately and efficiently as possible the flowering intensity and compare it to direct observations. Data were acquired on Malus domestica (apple) species, which exhibit differences in phenology and within-tree flower position that might necessitate adaptation for the assessment of their flowering intensity. We also conducted a comprehensive comparative analysis between two data acquisition protocols: aerial and terrestrial based stereo imagery providing top and lateral views of the trees respectively. They were tested on core-collections of apple and prunus located at INRAE experimental units in Montpellier and Avignon (France), respectively. For aerial based imagery, we employed photogrammetry to generate a colored point cloud. From the terrestrial based imagery, we extracted depth maps using stereovision. In both cases, we focused on the central canopy area to minimize interference with neighboring trees. Semantic segmentation was employed to identify pixels or points corresponding to each tree from the two types of data. We estimated the tree volume and performed a flower identification, using either color filters or semantic segmentation. Promising results were obtained by comparing the predictions from the pipeline with expert ground truth and heritability values. Subsequently, we examined the advantages and limitations of aerial and terrestrial based stereo imagery in assessing flowering intensity. This exploration presents interesting perspectives for high throughput phenotyping of flowering fruit tree collections of diversity.