利用可见-近红外光谱技术联合变量选择新方法对南丰蜜桔的可溶性固形物(SSC)进行快速无损检测研究, 以简化南丰蜜桔SSC预测模型和提高预测模型性能。 试验共采用300个南丰蜜桔样本, 校正集、 验证集及预测集样本分别为150, 75和75个。 采用QualitySpec型光谱仪在350~1 000 nm波段范围内采集样本光谱, 利用无信息变量消除(UVE)剔除无用信息波长变量, 再采用独立成分分析(ICA)提取光谱的独立成分, 最后应用最小二乘支持向量机(LS-SVM)建立南丰蜜桔的SSC预测模型, 并利用未参与建模的预测集样本对模型进行评价。 研究结果表明, 可见-近红外光谱技术联合UVE-ICA- LS-SVM对南丰蜜桔的SSC检测精度高。 UVE-ICA可以有效剔除无用信息波长变量, 提取特征光谱信息, 简化预测模型及提高预测模型性能。 UVE-ICA- LS-SVM所建立的南丰蜜桔SSC预测模型性能优于PLS, PCA-LS-SVM及ICA-LS-SVM预测模型, 其校正集、 验证集及预测集的决定系数和均方根误差分别为0.978, 0.230%, 0.965, 0.301%及0.967, 0.292%。

The objective of the present research was to assess soluble solids content (SSC) of Nanfeng mandarin by visible/near infrared (Vis/NIR) spectroscopy combined with new variable selection method, simplify prediction model and improve the performance of prediction model for SSC of Nanfeng mandarin. A total of 300 Nanfeng mandarin samples were used, the numbers of Nanfeng mandarin samples in calibration, validation and prediction sets were 150, 75 and 75, respectively. Vis/NIR spectra of Nanfeng mandarin samples were acquired by a QualitySpec spectrometer in the wavelength range of 350~1 000 nm. Uninformative variables elimination (UVE) was used to eliminate wavelength variables that had few information of SSC, then independent component analysis (ICA) was used to extract independent components (ICs) from spectra that eliminated uninformative wavelength variables. At last, least squares support vector machine (LS-SVM) was used to develop calibration models for SSC of Nanfeng mandarin using extracted ICs, and 75 prediction samples that had not been used for model development were used to evaluate the performance of SSC model of Nanfeng mandarin. The results indicate that Vis/NIR spectroscopy combined with UVE-ICA-LS-SVM is suitable for assessing SSC of Nanfeng mandarin, and the precision of prediction is high. UVE-ICA is an effective method to eliminate uninformative wavelength variables, extract important spectral information, simplify prediction model and improve the performance of prediction model. The SSC model developed by UVE-ICA-LS-SVM is superior to that developed by PLS, PCA-LS-SVM or ICA-LS-SVM, and the coefficient of determination and root mean square error in calibration, validation and prediction sets were 0.978, 0.230%, 0.965, 0.301% and 0.967, 0.292%, respectively.