Code, masked seed data and trained weights for: Cross-validation overestimates out-of-session accuracy in hyperspectral seed variety identification

This record accompanies the article "Cross-validation overestimates out-of-session accuracy in hyperspectral seed variety identification" by A. Uzdil and M. Koklu, submitted to Computers and Electronics in Agriculture. Single seeds of 17 seed lots of seven cereal and legume species, 16 registered cultivars and the grass pea population Siyah Tohum, were imaged with a pushbroom VNIR hyperspectral camera with 342 bands from 399.6 to 1003.2 nm at about 0.225 mm per pixel. The record holds 220 scans: 177 training trays with one cultivar per tray, 20 scans of grey-floor and 20 scans of black-floor mixed trays in which all 17 cultivars are placed in randomised layouts, and 3 hand-arranged trays. masked_data holds the seed masks, the calibrated reflectance spectra of all seed pixels with their coordinates, the seed and scan tables, the tray-level folds and the layout files. weights holds the trained weights of all models reported in the article. code holds the data preparation, model training and prediction code under the MIT licence. Four preview images show what the masked data contain. README.md describes every file, and CHECKSUMS.sha256 lists the SHA-256 sums. Caution: in the training trays, cultivar, seed lot, tray floor colour and acquisition session are confounded. Use the randomised mixed trays as the external test set. The raw hypercubes, about 244 GB, are available from the corresponding author on request.