提出一种基于流行降维的近红外光谱技术快速判别大米贮藏期的新方法。 采用近红外光谱仪获取陈年米和新米的反射光谱特征曲线, 利用直接正交信号矫正法(direct orthogonal signal correction, DOSC)对原始光谱进行预处理, 滤除光谱数据中与因变量Y矩阵无关的信号, 以消除无关信息对后续特征变量建模精度的影响。 采用Durbin-Watson和Run测试法定性分析光谱数据结构的非线性性, 并利用增强偏残差图(augmented partial residual plot)定量分析大米光谱曲线的非线性程度。 分别采用线性流行降维法包括主成分分析法(PCA)和多维尺度分析法(MDS)以及非线性流行降维法包括等距映射法(ISOMAP)、 局部线性嵌入法(LLE)和拉普拉斯特征映射法(LE)提取预处理后光谱数据的本征变量, 并结合核偏最小二乘方法(KPLS)建立本征变量与贮藏时间属性之间的耦合模型。 实验用陈年米和新米的样本数均为200个, 随机将训练集和测试集样本划分为300个和100个。 通过比较各个模型的预测结果得出, 基于ISOMAP非线性降维法提取的40个本征变量建立的回归模型预测效果最好, 预测相关系数(R2P)、 预测均方根误差(RMSEP)和预测相对分析误差值(RPD)分别为0.917, 0.187和2.698。 实验结果说明提出的方法对于大米贮藏期具有很好的鉴别能力, 该研究为今后大米贮藏期的快速无损检测提供了科学的手段。

This paper proposed a method for rapid identification of rice storage period based on manifold dimensionality reduction algorithms and near infrared spectroscopy (NIRS) technique. The reflection spectrum curve of old rice and new rice were obtained with a field spectroradiometer and the acquired spectral data was preprocessed with direct orthogonal signal correction method (DOSC) to filter the independent signal from the spectral data which is irrelevant with the dependent variable Y array and eliminate the influence and interference of the irrelevant information in the following chemometric analysis. The Durbin-Watson test and Run test methods were utilized to detect the nonlinearity which exists in the spectral data structure. The enhanced partial residual plot analysis method (Augmented partial residual plot) was employed to quantitative analysis of the degree of nonlinearity of the spectral data. Popular linear manifold dimensionality reduction methods including principal component analysis (PCA) method and multidimensional scaling analysis (MDS) method and popular nonlinear manifold dimensionality reduction methods including Isometries mapping method (ISOMAP), locally linear embedding (LLE) method and Laplacian Eigenmap method (LE) were used to extract the real variable from the preprocessed spectral data. Then, the intrinsic variable was taken as the input of the kernel partial least squares method (KPLS) to establish the relationship between the intrinsic variables and the storage time of rice samples. The number of experiment samples of the new rice and the old rice were 200 respectively and randomly separated into the training set with 300 samples and the test set with 100 samples. Through comparing the prediction results of the regression models which were established with different manifold reduction methods, the experiment results show that the prediction effects of the nonlinear-based models are superior to the linear-based models. Finally, the KPLS model established with 40 true variables extracted with ISOMAP approach achieved the optimal prediction effect. The prediction correlation coefficient (R2p), RMSEP (RMSEP) and relative prediction error value (RPD) were 0.917, 0.187 and 2.698, respectively. It was concluded that NIRS combined with ISOMAP-KPLS method can be successfully used to determine the storage period of rice accurately and quickly. The study provides a scientific means for rapid non-destructive detecting for rice storage period research in the future.