不同品种苹果糖度近红外光谱在线检测通用模型研究

Development of Multi-Cultivar Universal Model for Soluble Solid Content of Apple Online Using Near Infrared Spectroscopy

由于果实内部细胞结构、组成成分和光学传输特性的不同,品种差异会对近红外建模分析果实内部品质时产生较大的影响,以致原有模型无法高精度地预测果实品质参数。探讨开发不同品种近红外通用模型用于在线检测苹果内部品质的可行性。采用水果动态在线分选设备,设置运行参数为:积分时间100 ms,运动速度5 s^-1,采集包括冰糖心,红富士及水晶富士三个品种苹果的近红外漫透射光谱。分析了三个品种近红外漫透射光谱的响应特征,其光谱曲线走势基本一致,在650, 709和810 nm附近存在突出吸收峰,而在670, 750与830 nm附近存在波谷,其差异主要表现为光谱吸收强度的差异。采用多元散射校正, Savitzky-Golay卷积平滑及归一化处理方法,减少了不同品种引起的光谱信息差异。混合三个品种各校正集样本,采用偏最小二乘回归算法建立了不同品种糖度的通用模型,并利用无信息变量消除法(UVE)对建模变量进行筛选,最终得到的有效变量个数为155。所建立的UVE-PLS模型对验证集的决定系数,均方根误差以及残留预测偏差分别为0.80, 0.61%与2.21。在UVE筛选变量的基础上,采用连续投影算法再对建模变量进行选择,最终选出的变量个数为22。采用多元线性回归(MLR)方法建立了简化后的通用模型,对验证集的决定系数与均方根误差分别为0.78与0.64%。测试集用于评估最佳的不同品种糖度通用模型的实际性能,模型对每个品种测试集的潜变量数,决定系数与均方根误差分别为6~10, 0.77~0.79与0.45~0.75%。结果表明水果动态在线分选设备对不同品种苹果内部品质检测的潜力。通过建立通用模型,扩大了单一品种模型的预测范围,提高了模型在不同品种间的预测稳健性。并且采用合适的变量选择方法能够减少模型变量个数,降低模型复杂程度,并最终提高模型速率。开发不同品种水果内部品质通用模型在波长有限的近红外光谱设备中具有良好的潜在应用。

Cultivar variability influences the near-infrared modeling and analysis of the internal quality of the fruit due to the different cell structure, composition and optical transmission characteristics of the fruit so that the original model can not predict the fruit quality parameters with high precision. The feasibility of a multi-cultivar model’s development for the online determination of the internal quality of apple including "Candy Heart", "Red Fuji" and "Crystal Fuji" was investigated. Near infrared diffuse transmittance spectra of each cultivar were collected by the fruit sorting equipment under the condition of the interval time of 100 ms and motion speed of 5 s^-1. The spectral curves of all the cultivars were similar where the prominent absorption peaks were near 650, 709 and 810 nm, and troughs were near 670, 750 and 830 nm, and their variations were mainly reflected in the spectral absorption intensity. The spectral pre-process methods including multiplicative scatter correction, Savitzky-Golay smoothing and normalization were employed to filter out the variations in signals caused by the cultivars. Partial least squares regression method was used to establish the common model for the soluble solid content where the calibration sets of the total samples were combined. Uninformative variable elimination was used to select the modeling variables whose number of effective variables selected was 155, and the performance of the UVE-PLS model resulted in greater coefficient of determination for prediction of 0.80, lower root mean square error of 0.61% and higher residual prediction deviation of 2.21. Successive projections algorithm was employed to select the variables in the wavelengths selected by UVE and the number of variables selected was 22. Multivariable linear regression was used to establish the simplified model, which resulted in coefficient of determination for prediction of 0.78 and root mean square error of prediction of 0.64%. The test sets of all the cultivars were used to access the performance of best universal model, which resulted in latent variables of 6~10, coefficient of determination for prediction of 0.77~0.79 and root mean square error of 0.45%~0.75%. The results highlighted the potential of dynamic on-line sorting instruments for the testing of internal qualities of apples. The prediction range of the single cultivar model was expanded, and the robustness of prediction model among different cultivars were improved by establishing the common model. Appropriate variable selection methods can decrease the number of model variables, reduce the complexity of the model and ultimately increase the model rate. The development of the universal model of different cultivars for predicting internal quality has a good potential application in wavelength-limited near infrared spectroscopy equipment.