Multispectral Imaging Distinguished Seeds at Maturity Stages and Spikelet Grain Positions in Agropyron Cristatum Using Convolutional Neural Networks

Abstract

Effective screening and classification of forage seeds are becoming more and more important for seed industry and researchers. In this study, we focused on the impacts of three maturity stages (milk, dough, and ripe kernels) and two spikelet grain positions (spikelet basal position for SBP and spikelet upper position for SUP) on seeds of Agropyron cristatum, an excellent forage crop. Based on multispectral imaging (MSI) technique, we innovatively developed a pipeline combining normalized canonical discriminant analysis (nCDA) and convolutional neural network (CNN), nCDA-CNN, which showed the advances of spectral prediction. We compared its prediction performances with other six common machine learning models, including support vector machine (SVM), linear discrimination analysis (LDA), random forest (RF), extreme learning machine (ELM), nCDA, and CNN. This study showed that the accuracies of SVM, LDA, RF, ELM, and nCDA models in distinguishing fresh seeds at three maturity stages were 87.7%, 96.6%, 80.9%, 81.1%, and 91.1%, respectively, which were higher than those (73.8%, 64.9%, 73.0%, 74.4%, and 86.7%, respectively) in dry seeds. In contrast, nCDA-CNN received the best classification performance, with the highest accuracy of up to 100.0%. Furthermore, nCDA-CNN could predict SBP and SUP with the best accuracies of 80.0%-100.0%, while prediction accuracies of SVM, LDA, RF, ELM, nCDA, and CNN were only 72.5%-94.6%, 73.6%-97.6%, 66.6%-85.6%, 54.8%-81.0%, 66.7%-83.3%, and 47.5%-75.0%, respectively. In summary, we successfully developed a rapid and non-destructive approach to improve the identification of seeds with minor differences in A. cristatum, which could be implemented in extensive studies of seed examination in the future.

Keywords: seeds, maturity, seed spikelet grain position, multispectral imaging, Agropyron cristatum