Abstract
Accurate identification of crop varieties across growth stages is fundamental for material verification and trial management, providing a reliable basis for subsequent performance evaluation and elite accession selection in breeding programs. However, it remains challenging to differentiate intraspecific varieties due to subtle morphological variations among closely related accessions. Here, we present Img2Variety, a novel convolutional neural network (CNN)-based framework for crop accession identification from whole-plant images. Img2Variety builds on transfer learning by fine-tuning pre-trained CNNs. It is designed to adapt to plant datasets with a large number of accessions but limited samples per accession, thereby improving generalization across diverse accessions. To enrich feature diversity, we propose a novel growth stage and multi-view mixed augmentation (GMMA) strategy that leverages variation in viewing angles and developmental stages to promote feature learning. We also employ an adaptive cross-entropy (ACE) loss that emphasizes misclassified samples during training to improve identification performance. Img2Variety was evaluated using six CNN backbones on two datasets: one comprising 11,170 RGB images of 93 rice (Oryza sativa) accessions throughout the entire growth period, and another containing 5,599 RGB images of 224 maize (Zea mays) inbred lines across nine growth stages. Img2Variety achieved a peak accuracy of 88.66 % for rice and 79.95 % for maize, with an average relative improvement of 86.30 % over six baseline methods on the maize dataset. Notably, it exceeded 80.22 % accuracy for pre-heading rice and the maize tenth-leaf stage. These results highlight Img2Variety's effectiveness in crop variety identification and its potential for early-stage crop management. A web-based implementation is freely accessible at https://ngdc.cncb.ac.cn/opia/img2variety.