PLS和SMLR建模方法在水蜜桃糖度无损检测中的比较研究

在实际应用中，一些实验条件往往不能严格控制而存在变化，从而影响近红外光谱检测模型的稳健性。文章以50个常温和50个冷藏后的奉化水蜜桃样品组成温度混合样品集，经光谱杠杆值和狄克松检验法进行异常光谱剔除后，采用偏最小二乘法(PLS)和逐步多元线性回归(SMLR)对水蜜桃糖度进行建模分析。PLS的建模结果：校正集相关系数R_C=0.965, 校正均方根标准误差RMSEC=0.301°Brix，交叉验证R_CV=0.812，交叉验证均方根标准误差RMSECV=0.67°Brix，标准偏差与交叉验证均方根标准误差的比值RPD=1.72；SMLR的建模结果：校正集R_C=0.929，RMSEC=0.424°Brix, 交叉验证R_CV=0.887, RMSECV=0.532°Brix, RPD=2.16。SMLR的预测结果要优于PLS的预测结果，在SMLR分析中，在3个不同的光谱区域4 290～7 817，7 817～10 725，4 290～10 725 cm-1的RPD值分别为1.97，1.89，2.16。试验结果表明，将不同温度条件下的样品组成温度混合样品集，用PLS和SMLR建立的模型具有较好的预测效果。

Comparison of PLS and SMLR for Nondestructive Determination of Sugar Content in Honey Peach Using NIRS

Nondestructive fruit quality assessment in packing houses can be carried out using near infrared (NIR) spectroscopy. However, in industrial process, some experimental conditions (e.g. temperature, fruit variety) cannot be strictly controlled and their changes would reduce the robustness of the NIR-based models. In the present paper, a total of 100 honey fruits from two super markets were used as experimental materials. Fifty honey fruits were stored at room temperature and the other fifty samples were stored at 0-4 ℃. NIR diffuse reflectance spectra of the honey peaches were measured in the spectral range of 4 000-12 500 cm-1 using InGaAs detector. After outlier diagnosis using leverage values and Dixon test and spectra data pretreatment with Norris derivative filter (segment length: 5, gap: 5), partial least square (PLS) regression with standard normal variate (SNV) transformation and stepwise multilinear regression (SMLR) with multiplicative scatter correction (MSV) were used to establish calibration models based on first derivative spectra. Comparing the two calibration methods of PLS and SMLR, the performances of the models developed by SMLR were found much better than that by PLS method. The best results for PLS models were: correlation coefficient of calibration (R_C)=0.965, root mean square errors of calibration (RMSEC)=0.301° Brix，correlation coefficient of cross-validation (R_CV)=0.812, root mean square errors of cross-validation (RMSECV)=0.67° Brix and ratio of standard deviation to root mean square errors of cross-validation (RPD)=1.72, which were slightly worse than those for SMLR: R_C=0.929, RMSEC=0.424° Brix of calibration and R_CV=0.887, RMSECV=0.532° Brix of cross-validation and RPD=2.16. The RPD values for SMLR models in three different spectral regions 4 290-7 817, 7 817-10 725 and 4 290-10 725 cm-1 were: 1.97, 1.89 and 2.16, respectively. The performance of the model developed by SMLR in the 4 290-7 817 cm-1 region was much better than that in the 7 817-10 725 cm-1 region. The results indicated that the SMLR method could develop a good calibration model by selecting wavelengths insensitive to temperature and NIR spectra could be used for sugar content prediction of fruit samples with varied temperature when developing a global robust calibration model to cover the temperature range.