应用近红外漫透射光谱检测技术对不知火杂柑的可溶性固形物(SSC)进行在线检测具有十分重要的意义。 研究变量筛选方法对不知火杂柑可溶性固形物在线检测模型的影响, 为实现其快速、 准确的在线检测分级奠定基础。 实验把形状不整、 内藏瓤瓣的不知火杂柑作为研究对象, 选取560~930 nm的光谱, 采用偏最小二乘法(PLS)建立不知火杂柑可溶性固形物的在线检测模型, 并讨论不同的光谱预处理方法(卷积平滑(S-G)、 一阶微分(1st derivatives)等), 不同的变量筛选方法(移动窗口偏最小二乘法MWPLS、 遗传算法GA、 连续投影SPA)对PLS所建预测模型性能的影响。 经对比, 多元散射校正(MSC)能有效地消除光散射的影响, 遗传算法能大大地降低了建模的波长点数, 缩短了建模时间, 改善模型预测精度。 其最优PLS模型的RP=0.956, RMSEP=0.380, RC=0.967, RMSEC=0.340。 实验表明在线检测不知火杂柑的可溶性固形物是完全可行的。

It is of great importance to detect soluble solids content (SSC) of online testing in hybrid “Skiranui Tangerine” citrus by using near-infrared diffuse transmittance spectra. In order to lay a good foundation for accurate and rapid online classification, this study focuses on the influence of variable methods on soluble solids content in hybrid “Skiranui Tangerine” citrus. We selected the random shape hybrid “Skiranui Tangerine” citrus with segments inside as the research object. In spectral range of 560~930 nm, the calibration models were developed based on partial least squares (PLS) in this experiment. Firstly, different pretreatment methods such as Savitzky-Golay, the first derivative and so on were compared with PLS Modeling results. Then moving window partial least squares (MWPLS), genetic algorithm (GA) and successive projections algorithm (SPA) were employed to improve the predictive models. After comparing the results, light scattering can be effectively eliminated by the multiplicative scatter correction (MSC). Moreover, fewer variables and model optimization were carried out with GA. The best calibration model obtained with GA-PLS method had the correlation coefficient of prediction (RP) of 0.956, the root mean square errors of prediction (RMSEP) of 0.380, the correlation coefficient of calibration (RC) of 0.967 and the root mean square errors of calibration (RMSEC) of 0.340. The experiment showed that online detection of SSC of “Skiranui Tangerine” is completely feasible.