Fine-tuning improves deep learning-based digital phenotyping of lettuce genetic resources

Morphological characterization of agricultural genetic resources relies largely on visual assessment by experts, and applying the same criteria repeatedly to large collections is labor-intensive. Image-based deep learning can support this process, but systematic evidence on training strategies suited to genebank trait data remains limited.

This study targeted six morphological traits of lettuce genetic resources characterized according to the test guidelines of the International Union for the Protection of New Varieties of Plants (UPOV). Using genebank images accumulated from 2012 to 2025, six convolutional neural networks were trained with three training strategies under identical conditions, and their performance was compared.

Fine-tuning achieved the highest mean macro F1-score in 35 of the 36 backbone-trait combinations and significantly outperformed training from scratch and feature extraction, whereas no single backbone was superior across all traits. Misclassifications were concentrated in morphologically adjacent classes rather than randomly distributed, and for the ordinal traits, errors spanning two or more class levels were rare.

These patterns suggest that the remaining classification difficulty was associated in part with the morphological continuity between neighboring classes. The attainable performance also differed among traits, distinguishing traits for which the training strategy alone achieved relatively high performance from those for which it did not. These results provide empirical evidence for selecting training strategies and for determining the role of model-assisted characterization on a trait-by-trait basis in genebanks.