Hyperspectral imaging technology combined with deep learning for hybrid okra seed identification

Abstract

Crossbreeding is a plant breeding method of hybridizing the parents to form a different genetic variety to obtain anticipated excellent parental traits. The identification and screening of hybrid seeds are indispensable steps in the procedure of crossbreeding. In this study, deep learning (DL) and popular neural networks combined with widely used hyperspectral imaging (HSI) technology were applied to identify hybrid okra seeds. The hyperspectral images involving 18 okra varieties with a total of 18,931 seeds at a resolution of 5 nm were collected and processed in the spectral waveband of 948.17–1649.20 nm. Principle component analysis and linear discrimination analysis were applied to explore the visual clustering identification and the internal relation of the okra varieties. Four discriminant models, namely, extreme learning machine (ELM), back propagation neural network (BPNN), stacked sparse auto-encoder (SSAE), and convolutional neural network (CNN), were constructed and analyzed. Another primary research focused on the influence of the change in the input of the okra varieties on the models’ performance. Our results demonstrated that the CNN model reached the highest accuracy among the other models, and it had the strongest robustness. While the number of varieties grew continuously ranging from 7 to 18, all the models showed the decreasing classification accuracy. Moreover, the CNN achieved more stable results only falling by about 4.00% from 97.68% to 93.79%. The overall results indicated that the combination of DL, mining fully the information of spectral big data, and HSI could be a meaningful tool for identifying and detecting hybrid seed varieties. • Hyperspectral imaging combined with deep learning method produces a promising technique to identify hybrid okra seeds. • Deep learning showed the superior performance in hyperspectral images of different seeds classification. • This approach has potential for no-destructive seeds identification application.