基于高光谱散射图像的苹果压缩硬度和汁液含量无损检测

Non-destructive detection for compression hardness and juiciness in apple based on hyperspectral scattering image

压缩硬度和汁液含量是衡量苹果内部品质的两项重要指标。采用高光谱散射图像技术对苹果压缩硬度和汁液含量进行预测。已有研究表明,高光谱图像含有丰富的波谱信息,光谱值与测量值之间存在严重的非线性关系,简单的线性建模方法不能达到较高的预测精度。最小二乘支持向量机(Least Squares Support Vector Machine,LS_SVM)作为一种非线性建模工具,已用于解决小样本、非线性和高维数等实际问题。针对580个‘RedDelicious’苹果的高光谱散射图像,提取600～1000nm范围内的波谱信息,采用LS_SVM建立苹果的压缩硬度和汁液含量模型。研究结果表明,LS_SVM压缩硬度预测模型的相关系数为Rp=0.795,预测均方差为RMSEP=10.4KN/m,汁液含量的相关系数为Rp=0.568,预测均方差为RMSEP=1.20cm2,高于传统的偏微分最小二乘(PartialLeastSquares,PLS)建立的压缩硬度,模型精度Rp=0.744,RMSEP=11.4KN/m,汁液含量模型精度Rp=0.539,RMSEP=1.23cm2。

Compression hardness and juiciness were two important indexes in measuring apple internal qualities. Hyperspectral scattering of image technology to predict apple compression hardness and juiciness were introduced.Studies had shown that hyperspectral image contained rich spectrum information, and there was seriously nonlinear relationship between spectral value and measurements, so that simple linear modeling methods could not achieve high precision of prediction. Least squares support vector machine （LS_SVM）as a kind of nonlinear modeling tool, had used to solve the practical problems of small sample, nonlinear and high dimension. Five hundred and eighty ＇ Red Delicious＇ apples were used in the experiment. Hyperspectral scattering images were acquired in the range of 600-]000nm and the model of the apple compression hardness and juiciness were developed by using LS_SVM.The results showed that accuracies of prediction model for compression hardness and juiciness based on LS_SVM were better than that based on traditional partial differential least squares（PLS）. The correlation coefficient was 0.795 and the root meant that square error was ]0.4KN/m for compression hardness and the correlation coefficient was 0.568 and the root mean square error was ].20cm2 for juiciness by LS_SVM.The correlation coefficient was 0.744 and the root mean square error was ] ].4KN/m for compression hardness and the correlation coefficient was 0.539 and the root mean square error was 1.23cm2 for juiciness respectively by PLS. Key words, least squares support vector machine （LS_ SVM）; apple; compression hardness; juiciness; hyperspectral scattering image