为了提高对蓝莓果渣的开发利用, 探索了近红外光谱测定三种蓝莓(北陆、 蓝美1号、 灿烂)果渣中花色苷含量的可行性。 通过DA7200采集三种蓝莓果渣的近红外光谱, 利用PCA-MD对北陆、 蓝美1号、 灿烂果渣分别剔除1, 4和8个异常样本。 运用K-S划分样本集得到校正集(686个样本)和验证集(171个样本)。 对样本集分别进行归一化、 变量标准化(SNV)、 多元散射校正(MSC)、 Norris一阶导数(NFD)、 Norris二阶导数(NSD)、 SG卷积一阶导数(SGCFD)、 SG卷积二阶导数(SGCSD)、 Savitzky-Golay(SG)卷积平滑、 正交信号校正预处理, 并建立相应全谱PLS模型。 比较并选择MSC、 SGCSD、 SG卷积平滑、 正交信号校正, 进行预处理方法顺序组合的比较, 结果显示, 全谱PLS模型中最优预处理方法为正交信号校正+SGCSD+SG卷积平滑, 其R2c为0.940 0、 R2p为0.886 7、 RMSEC为0.722 5、 RMSECV为0.246 2、 RMSEP为1.000 5、 RPD为2.970 8。 利用SPA和CARS对预处理过的光谱数据分别进行波长变量的筛选, 依次建立PLS回归模型, 并定量分析其对蓝莓果渣花色苷的预测能力。 在所有预处理方法进行波长变量筛选中, SPA与CARS算法均可以有效地筛选出波长变量, 但SPA筛选出的波长变量, 无法全部建立PLS回归模型, 而CARS算法筛选出的波长变量, 均可建立PLS回归模型。 数据表明, CARS-PLS最佳组合为正交信号校正+MSC+SG卷积平滑+SGCSD, 选择波长数为25个, 相较于原始光谱, 其R2c从0.900 8增长到0.940 3, R2p从0.881 8增长到0.885 7, RMSEC从0.929 1减少到0.720 9, RMSECV从0.317 6减少到0.245 6, RMSEP从1.021 8减少到1.004 9, RPD从2.908 8增长到2.957 5。 近红外光谱的蓝莓果渣花色苷含量测定中, 正交信号校正表现出强大的去噪效果, CARS算法具有简化模型、 适用性较好和预测精度较高等优点。 研究结果表明, 应用近红外光谱技术可以较好地实现三种不同品种蓝莓果渣中花色苷含量的测定, 可为蓝莓果渣品质分级提供一种快速、 支持大样本量的检测方法。

To improve the development and utilization of blueberry pomace, the test measured the feasibility of near-infrared spectroscopy for the determination of anthocyanins in blueberry pomace of the three species which includes Northland, Bluebeauty No.1 and Brightwell. We gathered the near-infrared spectroscopy data of three blueberry pomaces through DA7200 and eliminated 1, 4 and 8 abnormal samples of Northland, Bluebeauty No. 1 and Brightwell respectively by principal component analysis-Mahalanobis distance. The K-S was used to divide the sample set into correction set (686 samples) and verification set (171 samples). Normalization, standardized normal variate (SNV), multivariate scattering correction (MSC), Norris first derivative (NFD), Norris second derivative (NSD), SG convolution first derivative (SGCFD), SG convolution second derivative (SGCSD), Savitzky-Golay (SG) convolution smoothing and orthogonal signal correction preprocess were performed on the sample set respectively, and the full spectrum PLS model was built accordingly. Preprocess methods with sequential combinations of MSC, SGCSD, SG convolution smoothing and orthogonal signal correction were compared. The results showed that the optimal preprocess method in the full spectrum PLS model was orthogonal signal correction+SGCSD+SG convolution smoothing, with R2c as 0.940 0, R2p as 0.886 7, RMSEC as 0.722 5, RMSECV as 0.246 2, RMSEP as 1.005, RPD as 2.970 8. Wavelength filtering algorithms SPA and CARS were used to screen the pre-processed spectral data. Then PLS regression model was established and the ability to predict anthocyanins in blueberry pomace was quantitatively analyzed. In the screening of wavelength variables for all pretreatment methods, both SPA and CARS algorithms can effectively screen out the wavelength variables, but the wavelength variables screened by SPA algorithm cannot be used to build PLS regression model, while the wavelength variables screened by CARS algorithm can. The data showed that the optimal combination of CARS-PLS was orthogonal signal correction+MSC+SG convolution smoothing+ SGCSD, with several selected 25 wavelengths. Compared with the original spectrum, its R2c increased from 0.900 8 to 0.940 3, R2p rose from 0.881 8 to 0.885 7, RMSEC decreased from 0.929 1 to 0.720 9, RMSECV dropped from 0.317 6 to 0.245 6, RMSEP changed from 1.021 8 to 1.004 9, and RPD was raised from 2.908 8 to 2.957 5. In the measurement of anthocyanin content in blueberry pomace by near infrared spectroscopy, the orthogonal signal correction has strong denoising effect, while CARS algorithm has the advantages of the simplified model, good applicability and high prediction accuracy. The result indicated that near-infrared spectroscopy could be used to determine anthocyanin content in blueberry pomace of three different varieties, and it can provide a fast and large sample size detection method for blueberry pomace quality classification.