基于高光谱图像技术的高粱品种识别研究  

Identification of Sorghum Breed by Hyperspectral Image Technology

作　　者：宋少忠 刘园园 周紫阳[3] 滕星[3] 李继红[3] 刘君玲[1] 高勋[2] SONG Shao-zhong;LIU Yuan-yuan;ZHOU Zi-yang;TENG Xing;LI Ji-hong;LIU Jun-ling;GAO Xun(School of Data Science and Artificial Intelligence,Jilin Normal University of Engineering and Technology,Changchun 130052,China;School of Physics,Changchun University of Science and Technology,Changchun 130022,China;Jilin Academy of Agricultural Sciences,Changchun 130033,China)

机构地区：[1]吉林工程技术师范学院数据科学与人工智能学院,吉林长春130052 [2]长春理工大学物理学院,吉林长春130022 [3]吉林省农业科学院,吉林长春130033

出　　处：《光谱学与光谱分析》2024年第5期1392-1397,共6页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(61575030);吉林省科技厅项目(20230202015NC,YDZJ202101ZYTS117);财政部-农业农村部-国家现代农业产业技术体系项目(CARS-06-14.5-A12)资助。

摘　　要：高粱是酿造白酒的重要原料,高粱内的成分对白酒中微量成分含量和品质十分重要,并且高粱品质影响着白酒的质量和风味,因此,无损快速鉴别高粱品种对于提高白酒质量是个迫切需要的重要问题。采用高光谱成像技术结合机器学习算法对高粱品种进行分类鉴别,通过高光谱成像技术,获取了10个品种高粱的高光谱谱线以及图像纹理数据。采用多元散射校正(MSC)对光谱进行预处理,并用连续投影算法(SPA)筛选出62个特征波段,采用灰度共生矩阵提取高粱的4种纹理特征,分别以高光谱数据和光谱-图像数据融合,采用PLS-DA、SVM、ELM和RF等4种机器学习算法模型对10个高粱品种进行分类识别。结果表明,高光谱经MSC预处理后,用SPA降维提取的高光谱特征波段可以代表全光谱的数据信息,提高了PLS-DA模型识别高粱品种的稳定性。10个品种高梁的分类准确度从67.58%提高到93.85%,识别精度提升了27%。高光谱数据与图像纹理特征数据融合后,PLS-DA基于模型全光谱和特征谱段的高粱品种分类识别精度分别提升到96.47%和97.16%,相比于单一的高光谱数据更适用于高粱品种分类识别。相比于SVM、ELM和RF三种分类机器学习算法模型结果,PLS-DA机器学习算法模型的高粱品种分类识别精度最好。研究证明了高光谱成像技术结合机器学习算法在高粱品种鉴别中的有效性,可实现快速精确的高粱品质检测。Sorghum is an important raw material for liquor brewing.The components of sorghum are very important to the trace components and quality of liquor,and the quality of sorghum can affect the quality and flavor of liquor.Therefore,the nondestructive and rapid identification of sorghum breeds is an urgent and important question for improving the quality of liquor.In this paper,hyperspectral imaging technology combined with a machine learning algorithm is used to classify and identify sorghum breeds.By using the hyperspectral imaging technology,hyperspectral spectral lines and image texture data of 10 breeds of sorghum are obtained at the same time.Multivariate scattering correction(MSC)is used to preprocess the hyperspectral spectrum,and a continuous projection algorithm(SPA)is used to screen 62 feature bands.The gray level co-occurrence matrix extracts four texture features of sorghum.The hyperspectral spectral data and spectral-image fusion data are used,respectively,and four machine learning algorithms,including PLS-DA,SVM,ELM and RF,are used to classify and identify the sorghum breed.The results show that the hyperspectral characteristic bands extracted by SPA dimensionality reduction can be represented by the data information of the full hyperspectral spectral information after MSC pretreatment,which improves the stability of the PLS-DA algorithm model in the identification of the sorghum breed.The identification accuracy of 10 breeds of sorghum is improved from 67.58%to 93.85%,and the identification accuracy is increased by 27%.After the fusion of hyperspectral spectral data and image texture feature data,the identification accuracy of the sorghum breed by using the PLS-DA model under the conditions of full-spectrum and feature spectrum is improved to 96.47%and 97.16%,respectively,which is more suitable for the classification and identification of sorghum breed compared with the single hyperspectral data.Compared with the results of SVM,ELM,and RF machine learning algorithms,the PLS-DA machine learning algorith

关 键 词：高粱 高光谱成像 机器学习算法 品种识别 

分 类 号：S323[农业科学—作物遗传育种]