3.3.2. Model for Total Anthocyanidins Content Prediction

The performances of total anthocyanidin content prediction models were parallel with the models for cyanidin content prediction because cyaniding was the vast majority of anthocyanidin in PCC. As shown in Table 3, the 1-VR value of the total anthocyanidins in the visible/NIR spectral bands from 400 to 1100 nm and 1100 to 2498 nm after no scattering processing and first-order derivative preprocessing was 0.944, the minimum SECV value was 685.994, and the external correlation coefficient RSQ was 0.968, which meant that the cross-test effect was very satisfying. Combined with the cross-validation result shown in Figure 2B, the line slope was 0.990, and the samples scattered with no big deviation. The content of total anthocyanidins in PCC can be accurately predicted using the model after pretreatment of no scattering processing and first-order derivative.

**Table 3.** Calibration equations of total anthocyanidins content in purple leaf Chinese cabbage using different pretreatment models.


<sup>1</sup> RSQ: external correlation coefficient; <sup>2</sup> SEC: standard error of calibration set; <sup>3</sup> 1-VR: internal cross-validation correlation coefficient; <sup>4</sup> SECV: standard error of cross-validation.

#### 3.3.3. External Validation of the Calibration Models

Using the mathematical model developed by WinISI III software, the samples not involved in the calibration were analyzed for external validation of the effectiveness of the developed model. The effectiveness of the validation was indicated by RSQ, SEP, and Bias. After validation, the RSQ were 0.947 and 0.951, respectively, for cyanindin and total anthocyanidins models at 400 to 1100 nm and 1100 to 2498 nm visible/NIR spectra, after no scattering processing and first-order derivative pretreatment (Figure 2C,D). The test deviation biases were small, which were −234.079 and −222.0, respectively. The slopes of the external validation prediction plots (the linear regression between the NIR predicted values and the chemically determined results) were 0.917 for cyanidin and 0.913 for total anthocyanidins. The validation samples were irregularly distributed on both sides of the line, and there was no big deviation, which meant that the models worked well; they could output accurate results for efficient and rapid screening of high anthocyanidins content materials. In addition, the validation results of delphinidin, pelargonidin, and peonidin prediction models showed very poor performance, as we could expect.
