Establishment of a UAV-based phenotyping method for European pear rust in fruit orchards

Plant phenotyping is still the bottleneck in genetic resource evaluation and fruit breeding. Therefore, this study aims to establish a UAV-based high-throughput digital phenotyping method for spatial detection using European pear rust as a model disease in the field. Over 800 training RGB images were acquired by low altitude drone flights and annotated using the Computer Vision Annotation Tool (CVAT). An initial image dataset of 188 images was used to train a standard YOLOv5 algorithm. As result 84% of the pear rust infected leaves and 85% of the healthy leaves were correctly detected. For subsequent quantification of disease symptoms, the ratio between diseased and healthy leaves in the image is determined. Accurate location of disease symptoms within the orchard is enabled by a novel photogrammetry approach on georeferenced image data.