GYMNSA dataset

Description

Annotated image dataset with different stages of European pear rust in orchards for UAV-based automatic symptom detection. The evaluation of fruit genetic resources regarding a resistance to pathogens is an essential basis for subsequent selection in fruit breeding. Both genetic analysis and phenotyping of defined traits are important tools and provide decision data in the evaluation process. However, the phenotyping of plants is often carried out "by hand" and remains the bottleneck in fruit breeding and fruit growing. The development of a digital and UAV (unmanned aerial vehicle)-based phenotyping method for the assessment of genotype-specific susceptibility or resistance against diseases in orchards would significantly increase the efficiency of plant breeding. In this framework, a workflow for drone-based monitoring of pathogens in orchards was developed using the European pear rust (Gymnosporangium sabinae) as model pathogen. We provide a dataset with expert-annotated high-resolution RGB images with pear rust symptoms. The UAV images present different pear genotypes, including varieties, wild species and progeny from breeding. The dataset contains manually labelled images with a size of 768 x 768 pixels of leaves infected with pear rust at different stages of development, labelled as class GYMNSA, as well as background images without symptoms. A total of 584 annotated images and 162 background images, organized into a training and validation set, are included in the GYMNSA dataset. This dataset can be used as a resource for researchers and developers working on drone-based plant disease monitoring systems.

Steps to reproduce

For data collection, ten UAV-flight campaigns were realized between 2021 and 2023 under various weather conditions and with different flight parameters in the experimental orchard of the Julius Kühn-Institute for Breeding Research on Fruit Crops in Dresden-Pillnitz (Germany). The raw UAV-RGB images were collected using the quadcopters DJI Phantom 4 Pro V2.0 (P4P) and DJI Matrice 300 RTK with Zenmuse P1 camera version 03.00.01.04 and 07.00.01.10 (2021). DJI GS Pro software version 2.0.16 (2021) and the DJI Pilot PE version 1.8.0 (2022) was used for the flight planning. The images were taken every 3 seconds during a flight speed of 1 meter per second. The flights were carried out with a front and side overlap of about 75 to 90 % at a flight altitude of approximately 5 to 12 meters above the ground. The average ground sampling distance (GSD) was 0.17 cm per pixel. The images were cropped into a size of 768 x 768 pixels without overlapping using the Python library Pillow version 9.5.0 by J. A. Clark (2015). The annotation tool CVAT version 1.1.0 by B. Sekachev et al. (2020) was used for the 2-point bounding box annotation. The image metadata was saved in an Excel spreadsheet using the software ExifTool version 12.3.8.0 by P. Harvey (2016). The annotation of the symptoms on the cropped UAV RGB images was done manually after checking the image quality. The labelling of pear leaves infected with pear rust involved five different stages of infection (early to late symptoms). All infection stages were labelled as "GYMNSA" (= infected). Both, whole infected leaves and overlapping infected leaves were labelled. Infected leaves of pear rust were also annotated if they showed strong light reflections in addition to the infection. Through 5-fold cross-validation with the software Scikit-Learn version 1.3.2 by F. Pedregosa et al. (2011) with the configuration shuffle = True and random_state = 42, the best division of the data set was selected and the annotated images and background images were divided into two folders for training and validation.

Institutions

Categories

Funding

Licence