基于中红外光谱和化学计量学算法鉴别核桃产地及品种

为探究中红外光谱快速检测核桃产地和品质的可行性，基于中红外光谱分析技术，并将化学计量学的算法应用于中红外光谱判别分析之中，对中国四大核桃主产区的10类主要核桃品种进行检测，取得较好效果。通过提取核桃粉末的光谱透射率，去除原始光谱首尾部分的明显噪声，对保留的700~3 450 cm-1范围的光谱采用小波分析（wavelet transform，WT）算法进行去噪预处理，并采用无信息变量消除结合连续投影算法（UVE-SPA）提取光谱特征波数，采用主成分分析法（PCA）对光谱定性分析，基于反向传播神经网络（BPNN）、极限学习机（ELM）、随机森林(RF)、径向基函数神经网络（RBFNN）及偏最小二乘判别分析（PLS-DA）对全谱和特征波数建模对比。在4类不同产地核桃判别中，得到12个特征波数：803，1 355，1 418，1 541，1 580，1 727，1 747，1 868，2 338，2 462，2 824和3 166 cm-1，基于特征波数分类的正确率高于全谱的分类结果，BPNN算法结合特征波数建模得到的识别正确率高达97%，RF算法分类判别效果最差，正确率仅69.70%；在10类不同品种判别中，得到10个特征波数：903，1 275，1 507，1 541，1 563，1 671，1 868，2 311，2 845和3 437 cm-1，基于特征波数分类的正确率依然高于全谱的分类结果，BPNN算法结合特征波数建模得到的识别正确率高达83.3%。在特征波数通用性方面，两组特征波数范围中有2个特征波数相同：1 541和1 868 cm-1，其他大多特征波数也都相近，将10类品种特征波数作为输入变量对4类不同产地的核桃进行分类，分类结果较差，因此，在10类品种监督值下选取的特征波数无法适用于4类产地的判别问题，由此推断，即使是同一原始数据，基于不同判别问题得到的特征波数在建模时通用性较差。结果表明，经UVE-SPA算法提取特征波数后，变量数可减少99%以上，有效地简化了模型，减少计算量，提高预测的稳定性；总体上，每个分类器的表现为：BPNN>RBFNN>ELM>PLS-DA>RF；基于小波变换结合特征波数选取和反向传播神经网络算法能有效地实现核桃的产地和品种识别。

Identification of Walnut Origins and Varieties with Mid-Infrared Spectroscopy Analysis Technique

To explore the feasibility of rapid detection of the origin and quality of walnut by using mid-infrared spectroscopy, mid-infrared spectroscopy and chemometrics algorithms were used to classify walnuts of ten varieties from four major origins and finally good results were achieved. After extracting the transmittance spectra of walnut powder, the apparent noise was removed in the head and the tail of the original spectrum, and the remaining spectrum of 700~3 450 cm-1 was denoised by wavelet transform (WT) algorithm. The spectral characteristic wavenumber was extracted by uninformative variable elimination combined withsuccessive projections algorithm (UVE-SPA). Qualitative analysis of the spectrum was performed by principal component analysis (PCA). Back propagation neural network (BPNN), extreme learning machine (ELM), random forests (RF), radial basis function neural network (RBFNN) and partial least squares discrimination analysis (PLS-DA) were used for modeling based on the full spectrum and characteristic wavenumbers. For the discrimination of four different origins, 12 characteristic wavenumbers were selected: 803, 1 355, 1 418, 1 541, 1 580, 1 727, 1 747，1 868, 2 338, 2 462, 2 824, and 3 166 cm-1, the discrimination accuracy of characteristic wavenumbers was much higher than that of full spectrum, and the accuracy of BPNN algorithm combined with characteristic wavenumbers reached 97%. The result of RF algorithm was the worst, and the accuracy was only 69.70%. For the discrimination of ten varieties, 10 characteristic wavenumbers were selected: 903, 1 275, 1 507, 1 541, 1 563, 1 671, 1 868, 2 311, 2 845, 3 437 cm-1, the discrimination accuracy of characteristic wavenumbers was still much higher than that of full spectrum. The accuracy of BPNN algorithm combined with characteristic wavenumbersreached 83.3%. In terms of the versatility of characteristic wavenumbers, there were two same characteristic wavenumbers in the two sets of characteristic wavenumbers: 1 541 and 1 868 cm-1, and most of the other characteristic wavenumbers were similar. The spectra based on characteristic wavenumbers of 10 varieties were used as input variables to discriminate walnuts’ origins, and the result was poor. Therefore, the characteristic wavenumbers selected under the supervisory value of 10 varieties could not be applied to discriminate 4 types of producing origins. Even with the same original data, characteristic wavenumbers selected based on different discriminant problems were less versatile in modeling. After extracting the characteristic wavenumbers by UVE-SPA algorithm, the discrimination results showed that the number of variables can be reduced by more than 99%, which effectively simplified the model, reduced the amount of calculation, and improved the stability of prediction. In general, the performance of each classifier is: BPNN>RBFNN>ELM>PLS-DA>RF. The experimental results showed that the identification of walnut origins and varieties can be realized effectively based on wavelet transform, characteristic wavenumber selection and back propagation neural network algorithm.