基于可见/近红外光谱和变量选择的脐橙可溶性固形物含量在线检测

为联合可见/近红外光谱技术和变量选择方法在线检测脐橙主要内部品质指标可溶性固形物(SSC),分别选定脐橙校正集和预测集样本141个和47个,脐橙运输速度为0.3m/s,利用USB4000微型光谱仪在线采集脐橙样本的可见/近红外光谱,先分别采用无信息变量消除(UVE)和遗传算法(GA)对650~950nm波段范围的波长变量进行预筛选,再分别利用竞争自适应重加权采样(CARS)及连续投影算法(SPA)对波长变量进一步筛选,并应用偏最小二乘(PLS)方法分别建立脐橙SSC的在线预测模型,并与原始光谱等建立的预测模型进行比较。结果表明,对于脐橙SSC,预筛选方法GA优于UVE方法,变量选择方法CARS优于SPA方法;GA-CARS及GA-SPA联合变量选择方法优于对应的单一变量选择方法CARS及SPA。在上述变量选择方法中,GA-CARS方法获得的结果最优,其所建立的脐橙SSC的PLS模型的校正集和预测集相关系数分别为0.933和0.824,校正集和预测集均方根误差分别为0.429%和0.670%,性能优于原始光谱建立的PLS模型,且建模波长变量数由1 385个下降为78个,仅占原波长变量数的5.63%。由此表明,GA-CARS联合变量选择方法可以有效筛选脐橙SSC的波长变量,提高预测模型的稳定性和预测精度。

Online detection of soluble solid content in navel orange based on visible/near infrared spectroscopy and variable selection

Soluble solids content (SSC) is the main internal quality index of navel orange, in order to detect the SSC of navel orange by the combination of visible/near infrared spectroscopy and variable selection method, 141 samples of calibration set and 47 samples of prediction set were used. The transportation speed of navel orange was 0.3 m/s. The visible/near infrared spectra of navel orange samples were collected online by a USB4000 micro spectrometer. Firstly, uninformative variable elimination (UVE) and genetic algorithm (GA) were used to prescreen the wavelength variables in the wavelength range of 650~950 nm, then competitive adaptive weighted sampling (CARS) and successive projections algorithm (SPA) were used to further screen the wavelength variables. Furthermore, partial least squares (PLS) method was used to establish the online prediction models of SSC of navel orange, and these prediction models were compared with the prediction model established using original spectra. The results indicate that,for SSC of navel orange, GA method is better than UVE method in pre screening, while CARS method is better than SPA method in variable selection. GA-CARS and GA-SPA combined variable selection method is better than the corresponding single variable selection methods CARS and SPA. GA-CARS method obtains the best results for SSC of navel orange among the above variable selection methods, with the correlation coefficients of PLS model of SSC of navel orange in calibration and prediction set of 0.933 and 0.824 respectively, and the root mean square errors of calibration and prediction set are 0.429% and 0.670%, respectively. The performance of GA-CARS-PLS model is better than that of PLS model established by original spectra, and the number of modeling wavelength variables reduces from 1 385 to 78, only accounting for 5.63% of the number of original wavelength variables. In conclusion, the combined variable selection method of GA-CARS can effectively screen the wavelength variables of SSC of navel orange, and improve the stability and prediction accuracy of the prediction model.