Data for: Genetic dissection of dehulling efficiency in sunflower

This dataset contains RGB images of dehulled sunflower (Helianthus annuus L.) achenes collected from line per se and hybrid trials conducted in 2023 and 2024 across six environments. The dataset was generated to facilitate the prediction of dehulling efficiency (DE03).
Achenes were dehulled twice using an RLSM dehuller. Images were taken under standardized conditions using a PENTAX K-7 digital camera. Samples were arranged on a grey plate (21.3 cm × 32.5 cm) to provide uniform background contrast. Between one and six images per sample were taken, with the sample rearranged between images to increase variability in orientation and spatial distribution.

The full dataset comprises 3,036 images representing 245 genotypes. For machine learning applications, the dataset was divided into three subsets:
• Training: 1,737 images
• Validation: 699 images
• Test: 600 images

Patches of the size 2048x2048 px² were extracted from the original images and rescaled by a factor of 2 to obtain a final size of 1024x1024 px². In the training and validation set, the amount of patches generated from one image differed to counter the long-tail distribution of different hull colors. For example, images of sunflower kernels with white hull were under-represented in the dataset, therefore more patches were extracted from each image.
In the test-set, 5 patches were extracted from each image.

The training and inference code for this data is available at: https://github.com/grimmlab/SunflowerDehullingEfficiency