Identification of hybrid rice strain based on near-infrared hyperspectral imaging technology
Liu Xiaodan,Feng Xuping,Liu Fei,He Yong
DOI: https://doi.org/10.11975/j.issn.1002-6819.2017.22.024
2017-01-01
Abstract:The selection and identification of seeds are a key link in the process of agricultural breeding. In this study, near infrared (874-1734 nm) hyperspectral imaging technology combined with chemometrics and image processing technology was successfully performed to identify and visualize strains of hybrid rice seeds. A total of 2700 samples of 3 different strains of rice seeds were collected, and all samples were divided into the calibration set and the prediction set according to the ratio of 2:1 using the SPXY algorithm. PCA (principle component analysis) was applied to explore the separability of different rice seeds based on the spectral characteristics of rice samples, and the preliminary results demonstrated that hybrid rice seeds of 3 different strains showed a trend of classification. The full spectrum has a large data volume, and contains a large amount of redundant and collinear information, which would affect the accuracy and calculation speed of the model. Since the optimal wavelength selection can help to extract important information from the whole data to improve the performance of the model while simplifying it, we adopted SPA (successive projections algorithm) to select sensitive wavelengths. Seven sensitive wavelengths (985.08, 1106, 1203.55, 1399.04, 1463.19, 1601.81, 1645.82 nm) were determined from the range of 975-1646 nm, and these wavelengths were related to functional groups in molecules (N-H, C-H, NH3+), which indicated the reliability of the selected wavelength for modeling. Partial least squares-discriminant analysis (PLS-DA) and support vector machine (SVM) were applied to build the classification models based on the full spectra and optimal wavelengths, and an excellent classification was achieved, with the classification accuracy of over 90% for all models. The SVM model performed better than PLS-DA, and especially the full spectrum-based SVM model achieved outstanding identification results, with 99.67% classification accuracy for calibration set and 97.11% for prediction set. Compared with full spectrum-based models, optimal wavelengths-based models performed relatively worse, but still offered correct discrimination rates of over 90.22%. This results revealed that the selected wavelength is effective and reliable, which can provide a reference for on-line discrimination of different strains of hybrid rice seeds. Combined with image processing technology, the visual prediction map could be generated by inputting the average spectra of each rice seed into the SPA-SVM model, and different colors would be employed to represent different kinds of seeds. It showed that the visual analysis of the sample could intuitively identify rice seeds of different strains by these methods. The overall results indicated that near infrared hyperspectral imaging technology can be used to identify and visually predict hybrid rice seeds. This research provides a new way for rapid screening and identification of seeds in the process of agricultural breeding.