烟草是我国重要的经济作物, 税收的重要来源, 为国家的经济发展做出了巨大贡献, 然而, 烟草病害严重影响烟叶产量与品质。 采用光谱分析技术对烟草病害进行早期防治具有非常重要的现实意义。 以接种烟草花叶病毒(TMV)与马铃薯Y病毒(PVY)的烟草为研究对象, 分别采集室内与室外培养的染病烟草叶片高光谱数据。 为实现对烟草病害的精准识别, 每隔两天对两种染病烟草进行光谱数据采集, 将每种病害数据详细地分成五个严重度等级, 最终获得1 697个在350~2 500 nm波段范围内的光谱数据。 为对烟草高光谱数据进行有效利用, 以支持向量机(SVM)为基础, 结合快速近邻波段选择算法(FNGBS)与归一化匹配滤波(NMFW), 提出一种聚类与排序相结合的波段选择算法(FNG-NMFW)。 FNG-NMFW首先采用FNGBS算法对烟草光谱进行精细分组, 再采用NMFW算法对各组波段进行排序以选择特征光谱, 实现烟草光谱特征提取与降维。 在波段选择的基础上, 采用SVM对烟草特征光谱进行分类, 最终实现高精度烟草病害检测。 研究结果显示: 该模型性能稳定, 在样本数量较少情况下, 即可实现TMV与PVY两种病害的高精度识别。 对于TMV1与TMV3, 该算法可以获得精度优于94%的检测结果, 对于PVY1与PVY3, 该算法精度接近90%, 表明该算法可有效完成两种病害早期的识别与预防工作。 与采用全波段光谱数据进行病害检测的模型相比, FNG-NMFW模型优势明显, 烟草病害检测结果总体精度达94.46%, 精度提高约1.5%, 检测时间由12.9 s缩短为1.1 s。

Tobacco is an important economic crop and source of tax revenue in our country. It makes a huge contribution to the country’s economic development. However, tobacco diseases affect the yield and quality of tobacco leaves seriously. Therefore, It is important that the use spectral analysis technology for early prevention and control of tobacco diseases. Objects of research are tobaccos inoculated with tobacco mosaic virus (TMV) and potato Y virus (PVY). The hyperspectral data of infected tobacco cultivated indoors and outdoors are collected respectively. In order to improve the detection accuracy of tobacco diseases, spectral data of two kinds of diseased tobacco are collected every two days, each disease data is divided into five severity levels in detail, and finally, 1 697 spectral data in the 350~2 500 nm band are obtained. In order to make effective use of hyperspectral tobacco data, this paper is based on a support vector machine (SVM), combined with a fast nearest neighbor band selection algorithm (FNGBS) and normalized matched filtering (NMFW), and proposes a combination of clustering and sorting Band selection algorithm (FNG-NMFW). Firstly, FNG-NMFW uses the FNGBS to group the tobacco spectrum finely and then sorts the groups of bands based on the NMFW algorithm to select the characteristic spectrum and realize the extraction and dimensionality of the tobacco spectrum. After completing the band selection, this paper uses SVM to classify tobacco characteristic spectra and achieves high-precision tobacco disease detection. The research results show that the model has stable performance and high accuracy. When the proportion of training samples is only 40%, an overall accuracy (OA) is better than 80%; when the number of feature bands is selected as 40, OA can be better than 85%. The algorithm can achieve higher accuracy for both TMV and PVY diseases, but the recognition accuracy of TMV is slightly lower than that of PVY. For the monitoring of TMV1 and TMV3, the algorithm can achieve monitoring with an accuracy better than 94%, and for the monitoring of PVY1 and PVY3, the accuracy of the algorithm is close to 90%, which shows that the algorithm can realize the early identification and prevention of two diseases. Compared with the model that uses full-band spectral data for disease detection, the FNG-NMFW model has obvious advantages. The accuracy of tobacco disease detection results is 94.46%, the accuracy is improved by more than 1.5%, and the modeling time is shortened from 12.9 seconds to 1.1 seconds.