为了探讨利用二维相关可见/近红外光谱法优选猪肉挥发性盐基氮(TVB-N)特征变量的可行性, 以贮藏时间为外扰, 研究了不同新鲜程度猪肉样本的二维相关光谱特性。 首先, 获取56个猪肉样本在贮藏1～14 d的400~1 000 nm范围的可见/近红外反射光谱, 经过标准正态变量变换(SNV)处理后, 基于全波段光谱建立TVB-N的偏最小二乘回归(PLSR)模型。 然后, 依据TVB-N实测值, 从中挑选出10个具有一定浓度梯度的样本(贮藏时间分别为0, 36, 72, 108, 144, 180, 216, 252, 288和324 h), 利用一阶导数对光谱进行预处理后, 根据不同样本之间的光谱差异, 选取7个波段用于二维相关光谱解析。 分析各个波段的二维相关同步谱和自相关谱, 从7个波段范围内共选取23个变量作为不同贮藏时间下与TVB-N相关的敏感波长, 并建立简化的PLSR模型。 相较于全波段光谱数据所建模型, 模型效果有所改善, 预测集决定系数R2p由0.792 1上升至0.865 8, 误差从3.658 2 mg·(100 g)-1下降至3.246 0 mg·(100 g)-1。 表明基于二维相关光谱对猪肉TVB-N特征变量进行优选的思路是可行的, 该方法能够从全光谱数据中筛选出与目标物质相关的敏感变量, 这也为近红外光谱特征波长选择提供了一个新的方法。

In order to investigate the feasibility of two-dimensional (2D) visible/near-infrared (Vis/NIR) spectroscopy method to optimize the characteristic variables of total volatile basic nitrogen (TVB-N) in pork, storage time was employed as external disturbance and 2D correlation spectral characteristics of pork samples with different freshness degrees were studied in this paper. First, Vis/NIR reflectance spectra in the spectral region of 400~1 000 nm of 56 pork samples stored for 1～14 days were collected. Partial least squares regression (PLSR) model was established to relate full-band spectra after pre-processed with standard normalized variate (SNV) and TVB-N values with determination coefficient in the prediction set (R2p) of 0.792 1 and standard error in the prediction set (SEP) of 3.658 2 mg·(100 g)-1. Then ten samples which had a certain concentration gradient were selected for 2D correlation spectrum analysis (with storage time of 0, 36, 72, 108, 144, 180, 216, 252, 288 and 324 h) according to the reference values of TVB-N determined by the standard methods. To eliminate the influence of noise and environmental temperature, the original spectra were pre-treated with first derivative and seven bands were selected for 2D correlation spectrum analysis according to the spectral differences between different samples. The wavelength ranges were 400~420, 450~465, 500~550, 555~580, 586~717, 726~787 and 860~960 nm, respectively. By analyzing synchronization spectrum and autocorrelation spectrum of each band, 23 variables were selected as the sensitive wavelengths to TVB-N. Then simplified PLSR model was built based on the selected feature variables. Compared with the model based on full band spectral data, the model performance was improved with R2p increased to 0.865 8 and the SEP dropped to 3.246 0 mg·(100 g)-1. The results showed that it was feasible to optimize the characteristic variables of TVB-N based on 2D correlation spectrum and this method was capable of selecting feature variables which were related to target attribute. The study also provided a new method to select the characteristic wavelengths from NIR spectra.