基于高光谱成像技术的西瓜种子活力等级分类方法研究

Research on the Classification Method of Watermelon Seed Vigor Level Based on Hyperspectral Imaging Technology

为了探讨水果种子不同活力等级的判别方法,以西瓜种子为研究对象,建立基于偏最小二乘判别(PLS-DA)、极限学习机(ELM)的高光谱图谱信息种子活力判别模型。首先,提取光谱信息,建立西瓜种子活力等级的光谱定性分析模型,结果表明,特征变量筛选方法UVE(无信息变量消除)结合建模方法PLS-DA得到的效果较好,分类正确率为100.00%,相关系数为0.86。其次,选取图像PC1权重系数,提取波长点为685、790、826、836、855 nm的特征图像,计算平均灰度值,建立基于图像特征的种子活力等级定性分析模型。结果表明,PLS-DA的误判率为6.67%,相关系数为0.85,优于ELM检测模型的误判率(10.00%)和相关系数(0.83)。高光谱成像技术的光谱和图像信息都能较好区分种子的活力等级,但基于光谱信息建立的判别模型优于基于图像特征建立的判别模型。

In order to explore the discrimination methods of different vigor levels of fruit seeds,hyperspectral mapping models based on partial least-squares discriminant analysis(PLS-DA)and extreme learning machine(ELM)were developed for vigor discrimination of watermelon seeds.Firstly,the spectral information was extracted to build a spectral qualitative analysis model of watermelon seed vigor class,and the results showed that the UVE characteristic variables screening method combined with the PLS-DA modeling method obtained better results,with a correct classification rate of 100.00%and a correlation coefficient of 0.86.Next,the image PC1 weighting coefficients were selected,the feature images with wavelength points of 685,790,826,836,855 nm were extracted,the average gray value was calculated,and the qualitative analysis model of seed viability level based on image features was established.The results showed that the false discrimination rate of PLS-DA was 6.67%and the correlation coefficient was 0.85,which was better than the false discrimination rate of ELM detection model(10.00%)and the correlation coefficient(0.83).Both hyperspectral spectrum and image information could better distinguish seed vigor classes,but the discriminative model built by spectral information was better than the discriminative model built by image features.