**4. Discussion**

As an osmoregulatory substance, anthocyanins are one of the most important pigments in plant leaves. It has an irreplaceable role in improving the cold, drought, and disease resistance of plants, and, therefore, monitoring the content of anthocyanidins in plants can help to understand the physiological state of plants [3]. Meanwhile, the benefits for human health of anthocyanidins have drawn a great deal of people's attention. At present, the detection of anthocyanidins contents in plants and plant products mainly uses HPLC or HPLC-MS method; the application of rapid and non-destructive detection using NIR is still in its initial stage, but high throughput, convenient operation, and no use of organic solvents will make NIR a powerful support tool in horticulture practice and agri-food industry. Huang et al. [25] proposed a NIR spectroscopic detection method based on an ant colony algorithm (ACO) combined with interval partial least squares (iPLS) in order to detect anthocyanidins content in flower tea quickly and accurately, indicating that NIR spectroscopy has promising applications in measuring total anthocyanidins in plants. NIR spectroscopy can be used to determine the anthocyanidin content of berries in completely satisfied results without breaking the composition of the berries [26]. In this study, we successfully developed suitable prediction models for cyanidin and total anthocyanin content in PCC, and they could be applied in the breeding practice of PCC to realize rapid and efficient screening of high-quality breeding materials.

NIR spectroscopy belongs to an indirect analysis technology; the accuracy of the prediction result relies on the quality of the calibration models. So, the establishment of a high-quality model, with accurate chemical analysis and spectrum scanning data, strong anti-interference capability, and broad enough representation, is vitally important. A large number and representative sample sets are essential factors for model building. Additionally, an appropriate algorithm to divide sample subsets is also critical [27]. In order to expand the application scope of our established models, further improvement using a larger number of PCC samples with different breeding backgrounds and distinct phenotypes is required. In terms of algorithms for NIR model establishment, there are several regression methods frequently used for the prediction/quantification of chemical content, including multiple linear regression (MLR), principal component analysis (PCA) for the exploration of the data, and partial least squares regression (PLSR) analysis to obtain a quantitative prediction of the parameters of interest [12,28]. Among them, PLSR is the most widely used multivariate statistical data analysis method for quantitative analysis of the NIR spectrum, with strong anti-interference ability. In this study, we used a PLSR method to process the spectral data, and the quantitative prediction results were satisfied.

Compared to chemical analysis methods, the sensitivity of NIR spectroscopy is relatively low and cannot be used for trace analysis, but its modeling is suitable for the detection of components with high content and a wide range of variation. In a previous report, a satisfied NIR prediction model was established to detect anthocyanidin content in flower teas with a content range of 0.17 to 1.60 mg/g [25]. In this study, the model prediction performance of cyanidin content and total anthocyanidin content with a wide range of variation was relatively good, which could be used for rapid screening of breeding materials and prediction of anthocyanidin content in PCC breeding practice. Meanwhile, the prediction model performance of delphinidin, pelargonidin, and petunidin with less abundant contents in PCC was very poor. Considering the contents of these three anthocyanidins were relatively low, their contribution to the phenotype and nutritional quality of PCC could be neglected. However, in other plant materials which contain a much higher proportion of these anthocyanidins, much more samples with a wide range of contents need to be included, and further optimization of their model-building methods is needed.

There are still some parts of NIR spectroscopy detection technology that need to be improved, but with the development of algorithm, spectroscopy, and artificial intelligence, the predictive ability, accuracy, and operability of this technology will continue to be improved on the original basis. With its obvious time-saving, high throughput, and nondestructive advantages, NIR spectroscopy will certainly have a broader development prospect in the agricultural, food industry, and market inspection.
