基于高光谱的马铃薯微型种薯分类检测

Classification Detection of Potato Micro Seed Potato Based on Hyperspectral

采用高光谱分析技术结合模式识别,建立了8种马铃薯微型种薯(大西洋、荷兰-14、荷兰十五041、荷兰十五Q8、冀张薯12号、冀张薯8号、兴佳2号和Y2)的分类检测方法。采集276个种薯样本,对860~1700 nm的原始光谱进行标准化、11点Savitzky-Golay平滑和4点差分一阶导数光谱预处理,将预处理后的光谱数据进行主成分分析,发现前3个主成分的累积贡献率为95.12%,包含了原始光谱的大部分信息,可作为分类变量。再分别使用线性判别分析、BP神经网络和支持向量机进行分类建模。最终通过分层、分步骤建立了8种马铃薯微型种薯的分类模型。首先采用线性判别分析模型区分大西洋、荷兰-14、荷兰十五041和其它品种,平均正确识别率达88.79%。再建立BP神经网络模型将其它品种样本区分为两类,一类为冀张薯8号和Y2,另一类为荷兰十五Q8、冀张薯12号和兴佳2号,平均正确识别率达93.24%。最后以BP神经网络模型区分冀张薯8号和Y2,平均正确识别率为77.78%;以支持向量机分类模型区分荷兰十五Q8、冀张薯12号和兴佳2号,平均正确识别率为87.23%。该研究建立的8种马铃薯种薯分步骤、分层分类识别模型的平均正确识别率达89.75%,表明高光谱光谱分析技术可用于马铃薯微型种薯的分类检测。

Using hyperspectral analysis technology combined with pattern recognition,the classification and detection methods of eight potato micro seed potatoes(Daxiyang,Holland-14,Holland fifteen 041,Holland fifteen Q8,Jizhangshu 12,Jizhangshu 8,Xingjia 2 and Y2)were established.276 seed tuber samples were collected.The original spectra of 860-1700 nm were preprocessed by standardize,11 points Savitzky-Golay smoothing and 4 points differential first derivative.Principal component analysis showed that the cumulative contribution rate of the first three principal components was 95.12%,including most information of the original spectra,and could be used as classification variables.Then,linear discriminant analysis,BP neural network and support vector machine were used for classification modeling.Finally,the classification models of 8 potato micro seed potatos were established by stratification and step by step.Firstly,the linear discriminant analysis model was used to distinguish Daxiyang,Holland-14,Holland fifteen 041 and other varieties.The average correct recognition rate was 88.79%.Then BP neural network model was established to divide the samples of other varieties into two categories:Jizhangshu 8,Y2,and Holland fifteen Q8,Jizhangshu 12,Xingjia 2,with an average correct recognition rate of 93.24%.Finally,the BP neural network model was used to distinguish Jizhangshu 8 and Y2,with the average correct recognition rate of 77.78%;and the support vector machine classification model was used to distinguish Holland fifteen Q8,Jizhangshu 12 and Xingjia 2,with the average correcct recognition rate of 87.23%.The method was applied to the classification detection of eight potato seed potatos with the average correct recognition rate of 89.75%,which indicated that the hyperspectral analysis technology could be used for the classification and detection of potato micro seed potatos.