近红外高光谱的脐橙粒化检测研究

Detection of Citrus Granulation Based on Near-Infrared Hyperspectral Data

脐橙粒化影响消费者食用口感,降低品质,受到广大果农和消费者的关注。脐橙粒化的检测是一项具有挑战性的任务,对品质分级具有重大意义。以不同粒化程度的赣南脐橙为研究对象,探究利用高光谱检测实现对赣南脐橙粒化程度定性判别的可行性。肉眼是无法判断脐橙粒化程度的,因此对脐橙样本做好序号标记后先测光谱再切开判断粒化程度,按照粒化程度分为无粒化(粒化面积为0%)、轻度粒化(粒化面积小于25%)、中度粒化(粒化面积25%~50%),每类各58个脐橙样品。在这三类脐橙底部均匀取3个点,每类174个样本,共计522个样本数据用作构建原始光谱矩阵。利用近红外高光谱成像系统采集样本397.5~1014 nm波段内的高光谱图像信息,再利用ENVI4.5软件通过选择感兴趣区域(ROI)提取样本的平均光谱信息。采用主成分分析(PCA)、连续投影算法(SPA)、无信息变量消除(UVE)三种降维方法对光谱数据进行降维处理,消除无关变量,提取有用信息。原始光谱176个波长,PCA挑选出6个主成分因子,SPA挑选17个特征波长,UVE挑选54个特征波长。以全谱数据和三种降维方法挑选出来的变量作为输入分别建立偏最小二乘判别分析(PLS-DA)和最小二乘支持向量机(LS-SVM)模型。建立的PLS-DA建模方法,PCA-PLS-DA误判率最高为25.58%,UVE-PLS-DA误判率最低为5.38%。基于RBF-Kernel和LIN-Kernel两种核函数下的LS-SVM建模方法,整体上RBF-Kernel建模效果优于LIN-Kernel,UVE波长筛选后建立的模型效果优于其他降维方法且降低了模型的误判率。基于RBF-Kernel的UVE-LS-SVM模型效果最佳,检测精度最高,分类总误判率为0.78%,达到最佳效果。该研究结果表明建立的模型能很好地对不同粒化程度的脐橙进行判别,该模型仅采用30.68%的数据,在降低光谱空间维度的同时还降低了误判率,对促进脐橙产业的品质分级发展具有一定的现实意义。

The granulation of navel orange affects consumers’taste and reduces its quality.It has attracted the attention of fruit farmers and consumers.The detection of navel orange granulation is challenging and has great significance for quality classification.In this paper,the different granulation degrees of Gannan navel oranges are used as the research object to explore the qualitative determination of the granulation degree of Gannan navel oranges by using hyperspectral detection technology.Since the degree of granulation of navel oranges cannot be judged by the naked eye,the samples of navel oranges are marked with serial numbers,and then the spectrum is measured.Finally,the samples were cut to determine the degree of granulation.According to the degree of granulation,it is classified as non-granulation(the granulation area is 0%);light granulation(granulation area less than 25%);and medium granulation(granulation area 25%~50%).Take 3 points uniformly at the bottom of these three types of navel oranges,each with 174 samples,and a total of 522 sample data are used as the rows for constructing the spectral matrix.The near-infrared hyperspectral imaging system was used to collect the hyperspectral image information of the sample in the 397.5~1014 nm band and then use the ENVI 4.5 software was used to extract the average spectral information the sample by selecting the Region of Interest(ROI).Three dimensionality reduction methods:Principal Component Analysis(PCA),Successive Projections Algorithm(SPA),and Uninformative Variable Elimination(UVE)are used to reduce the dimensionality of the spectral data to eliminate irrelevant variables and extract useful information.The original spectrum has 176 wavelengths.PCA selects 6 principal component factors.SPA selects 17 characteristic wavelengths,and UVE selects 54 characteristic wavelengths.The full spectrum data and the variables selected by the three-dimensionality reduction methods are used as input to establish Partial Least Squares Discriminant Analysis(PLS-DA)and Least Squares Support Vector Machines(LS-SVM)model.In the established PLS-DA modeling method,the highest false positive rate of PCA-PLS-DA is 25.58%,and the lowest false-positive rate of UVE-PLS-DA is 5.38%.The LS-SVM modeling method is based on the two kernel functions of RBF-Kernel and LIN-Kernel,and the effect of RBF-Kernel modeling is better than that of LIN-Kernel generally.And the model established after UVE wavelength screening is better than other dimensionality reduction methods,which reduces the model’s false positive rate.The UVE-LS-SVM model based on RBF-Kernel has the best effect and the highest detection accuracy,and the total misjudgment rate of classification is 0.78%,achieves the best results.This study shows that the established model can distinguish navel oranges with different granulation degrees.The model reduces the spectral dimension while also reducing the misjudgment rate with only 30.68%of the data,which is useful for promoting the quality of the navel orange industry with certain practical significance.