Résumé
Fruit tree flowering and reproduction processes are crucial for production. Flowering time is currently impacted by global warming affecting orchard production. The screening of genetic resources could provide new keys for fruit tree adaptation. Proxy-detection has great potential to help characterize plant cultivars and select the best performing ones. In this context, the ambition of this project is to develop new methodologies for predicting flowering dates using aerial images at regular time intervals. For this, we used a core-collection of 241 apple French cultivars with four repetitions per cultivar, planted in 2014 in Montpellier (south of France). RGB and multispectral images were acquired on the orchard, using drone-borne sensors, resulting in orthomosaics of the entire orchard. For each type of sensor, four periods of acquisition in 2021 (between June and November) were used to characterize different foliage phenological stages. In parallel, expert notations of the flowering dates were carried out in the following spring (2022). The eight orthomosaics were split into patches representing individual trees. Different methods of machine and deep learning were tested to predict the flowering dates from the patches. First, NDVI and NDRE were determined from each patch and regression methods, such as Lasso, Random Forest, Gradient Boosting, were applied. A convolutional neural network (CNN), inspired from U-Net, was then tested. This CNN was tested on individual patches and on concatenated patches of the different months. Finally, a clustering method was used to identify flowering periods by grouping trees with similar flowering dates and a CNN was trained to predict the flowering period of each tree. This last approach proved to be the most accurate with an RMSE of approximately 10 days. This work illustrates the interest and limitation of deep learning on images for the phenotyping of traits representing genotypic behavior at different tree developmental stages.