Maize Seed Variety Identification Using Hyperspectral Imaging and Self-Supervised Learning: A Two-Stage Training Approach Without Spectral Preprocessing

Rapid and non-destructive variety identification is essential for screening maize seeds for different end-uses such as food, feed, and breeding. Hyperspectral imaging (HSI) is one of the most commonly used techniques in such seed classification. Typically, after acquiring hyperspectral images of seeds, the spectral domain signals need to be preprocessed and a classifier need to be designed. The traditional method is to find a appropriate spectral preprocessing method through trial-and-error experiment, which is time-consuming, laborious and has high risk of misuse preprocessing. In view of this, this paper proposes a self-supervised learning method that includes pre-training and fine-tuning phases. In the pre-training phase, a model was trained on the unlabeled raw spectral data in an unsupervised manner to obtain general representations. In the fine-tuning phase, the pre-trained model was fine-tuned with the goal of the seed classification task and trained in a supervised manner on labeled spectral data. Experimental results showed that the proposed method did not rely on spectral preprocessing, and its performance was superior to other existing seed classification methods. In addition, the self-supervised pre-trained model significantly outperformed the non-pre-trained model in the downstream seed classification task, and obtained good generalization ability. Overall, this method combined with HSI for seed quality evaluation has broad application prospects.