UDATNN: A modeling scheme integrating unsupervised domain adversarial learning and tri-training strategy for variety recognition of maize seeds with domain shift

Abstract

Constructing an accurate classification model is the key to realizing maize seed varieties identification based on optical sensing techniques. However, the optical information (features) of seeds collected by optical sensing technology is easily affected by the planting environment (e.g., origin, year), which makes the training samples used to construct the model (source domain, SD) and the samples to be recognized (target domain, TD) subject to domain shift (DS), and ultimately undermines the recognition performance of the model. In this study, a modeling scheme, called unsupervised domain adversarial tri-training of neural networks (UDATNN), is proposed for maize variety recognition in the presence of DS. Firstly, an unsupervised domain adversarial learning approach is used to map the raw features of the SD and TD into a low-dimensional feature space in order to achieve feature alignment of the SD and TD, and to improve the feature discriminability of different varieties of seeds. Subsequently, the aligned low-dimensional features are used as inputs of classifiers (random forest) and part of the target domain samples are iteratively selected and given pseudo-labels according to the tri-training strategy. Finally, these samples assigned with pseudo-labels (called updating samples) together with the training samples in SD are used to re-train the classification model constructed based on the unsupervised domain adversarial strategy, for improving recognition accuracy of seed varieties with DS scenario. Hyperspectral images of a total of 4584 maize seeds, including 7 varieties (each variety produced from two or three years) were collected to verify the performance of the proposed scheme. The recognition accuracy of the target domain reaches 91.5%, 94.1%, and 92.6% under 3 different domain shifts, which improves the recognition performance nearly by 8%–22% compared with the no-transfer model and the traditional transfer model. The model performance was further discussed by calculating precision, recall, and F1 score to achieve satisfactory results. The robustness of the model was also verified by discussing the randomness of the update samples and the effect of the number of samples in the source domain on the model performance through 100 random experiments and multiple experimental comparisons. The proposed UDATNN scheme can be used as a new framework to address the nondestructive identification of seed varieties under domain shift conditions.