大米拉曼光谱不同预处理方法的相近产地鉴别研究

用相近产地的大米代替独有的地理因素形成的地域品牌大米，消费者难以辨别。基于拉曼光谱技术，试验对比不同预处理方法包括一阶导数+平移平滑、二阶导数+平移平滑，小波变换+去除基线三种常用的预处理方法，另外提出一种改进的分段多项式拟合+去除基线共四种预处理方法，分别结合偏最小二乘法实现相近产地大米的鉴别分析，提出一种最佳的鉴别相近产地大米的预处理方法。首先用拉曼光谱仪采集了黑龙江省依安县3个相近产地大米的150个拉曼位移为200～3 100 cm-1的大米光谱样本，再对原始拉曼光谱分别用一阶导数+平移平滑、二阶导数+平移平滑、小波变换+去除基线、分段多项式拟合+去除基线进行光谱预处理。分别从每个产地选取33个样本进行训练，并对未知的51个样本建立了基于偏最小二乘法的鉴别分析模型，在训练集中一阶导数+平移平滑的预处理方法相关系数值最大、均方误差和均方根误差最小，小波变换+去除基线的预处理方法相关系数值最小、均方误差和均方根误差最大；在测试集中采用3点2次拟合+去除基线的预处理方法的相关系数值最大、均方误差和均方根误差最小，二阶导数+平移平滑的预处理方法最差。最后再通过PLS建模结果得知，在训练集中，采用四类九种预处理的方法对三个产地大米的总识别率均为100%；在测试集中，采用3点2次拟合+去除基线对三个产地大米总识别率为100%，采用5点2次拟合+去除基线对三个产地大米总识别率为52.9%，其他分段多项式拟合介于二者之间；采用一阶导数+平移平滑、二阶导数+平移平滑和小波变换的总识别率分别为88.2%，86.2%和96.1%；从中发现，分段式多项式拟合中的3点2次拟合+去除基线的优势明显，与其相关系数、均方误差、均方根误差结果吻合，总体识别率高，鉴别效果稳定。

Identification of Rice From Similar Areas With Different Pretreatment Methods of Raman Spectrum

It is difficult for consumers to distinguish regional rice brands formed by geographical factors instead of rice of similar producing areas.Based on Raman spectroscopy,the experiment compared three common pretreatment methods including first derivative+translational smoothing,second derivative+translational smoothing,wavelet transform+baseline removal.In addition,an improved piecewise polynomial fitting+baseline removal was proposed,and a total of four pretreatment methods respectively were combined with partial least square method to identify rice of similar origin,and an optimal pretreatment method for identifying rice of similar origin was proposed.Firstly,150 rice spectral samples with a Raman displacement of 200~3100 cm^-1 were collected by Raman spectrometer from three similar producing areas in an county,Heilongjiang province.Then,the original Raman spectra were preprocessed by first derivative+translational smoothing,second derivative+translational smoothing,wavelet transform+baseline removal,and piecewise polynomial fitting+baseline removal.A total of 99 samples were selected from 33 samples from each origin for training,and a partial least square analysis model based on partial least square method was established for the unknown 51 samples.In the training set,the preprocessing method of first derivative+translation smoothing had the maximum correlation coefficient value,the minimum mean square error and the minimum root mean square error.Andwavelet transform+baseline removal hadthe minimum correlation coefficient value,the maximum mean square error and the maximum root mean square error.In the test set,the preprocessing method of 3 points and 2 times fitting+baseline removal had the maximum value of correlation coefficient,the minimum mean square error and the minimum root mean square error,and the preprocessing method of the second derivative+translation smoothing was the worst.Finally,the PLS modeling results showed that,in the training set,the correct discrimination rate of rice from three producing areas was 100%by using four kinds and nine pretreatment methods.In the test set,the total recognition rate of rice from three producing areas was 100%by using 3-point 2-time fitting+baseline removal,and 52.9%by using 5-point 2-time fitting+removing baseline.Other piecewise polynomial fitting was between the two.The total recognition rates of the first derivative+translation smoothing,the second derivative+translation smoothing and the wavelet transform were 88.2%,86.2%and 96.1%,respectively.It is found that the preprocessing method of 3-point 2-time fitting+removing baseline in piecewise polynomial fitting has obvious advantages and is consistent with the results of the correlation coefficient,mean square error and root mean square error,with high overall recognition rate and stable identification effect.