摘要
纹枯病是水稻生产中三大病害之一,其早期检测对病害的及时防控、保证粮食安全具有重要意义。高光谱技术为水稻病虫害高通量、实时监测提供了有效的技术手段。基于高光谱病害检测中高光谱的降维,或检测特征的提取至关重要,利用2017和2018两年水稻盆栽纹枯病接种试验与大田纹枯病调查试验样本高光谱数据,探讨了分窗Gram-Schmidt变换的高光谱数据降维与特征波段提取,构建纹枯病检测模型,对比分析了本研究方法与主成分分析、连续投影法高光谱降维效果与病害检测精度。结果表明:基于分窗GramSchmidt变换可有效实现高光谱数据的降维,盆栽样本高光谱降到4维,纹枯病检测模型决定系数R2为0.8373,均方误差MSE为0.0406;大田样本高光谱降到4维,纹枯病检测模型决定系数R2为0.9701,均方误差MSE为0.0065。主成分分析法降维,盆栽样本高光谱降到6维,纹枯病检测模型决定系数R2为0.7931,均方误差MSE为0.049,大田样本高光谱降到6维,纹枯病检测模型决定系数R2为0.9658,均方误差MSE为0.0078;连续投影法降维,盆栽样本高光谱降到8维,纹枯病检测模型决定系数R2为0.8132,均方误差MSE为0.0466,大田样本高光谱降到4维,纹枯病检测模型决定系数R2为0.9685,均方误差MSE为0.0072。对比主成分分析法和连续投影法,基于分窗Gram-Schmidt变换的高光谱降维效果与纹枯病检测精度均效果较好,可为高光谱降维与水稻纹枯病防治提供一定的理论基础和技术支撑。
Sheath blight is one of the three major diseases in rice production. Early detection of sheath blight is of great significance for timely prevention and control of diseases and ensuring food security. Hyperspectral technology provides an effective technical means for high-throughput and real-time monitoring of rice pests and diseases. Based on Hyperspectral dimensionality reduction or detection feature extraction in hyperspectral disease detection, this study used hyperspectral data of rice potted sheath blight inoculation test and field sheath blight investigation test samples in 2017 and 2018, discussed the dimensionality reduction and feature band extraction of hyperspectral data by window Gram-Schmidt transform, constructed the detection model of sheath blight, and compared and analyzed the research methods. The results showed that the dimensionality of hyperspectral data could be reduced effectively based on split-window Gram-Schmidt transform. The hyperspectral data of potted samples could be reduced to 4 dimensions. The determination coefficient R2 of sheath blight detection model was 0.8373 and the mean square error MSE was 0.0406. The hyperspectral data of field samples could be reduced to 4 dimensions. The determination coefficient R2 of sheath blight detection model was 0.9701 and the mean square error MSE was 0.0065. Principal Component Analysis(PCA) reduced the dimension of potted samples to 6 dimensions, the determination coefficient R2 of sheath blight detection model was 0.7931, the mean square error MSE was 0.049, the hyperspectral of field samples was reduced to 6 dimensions, the determination coefficient R2 of sheath blight detection model was 0.9658, and the mean square error MSE was0.0078;the successive projections algorithm reduced the dimension of potted samples to 8 dimensions, and the determination coefficient R2 of sheath blight detection model was 0.8132, respectively. The mean square error MSE was 0.0466, the hyperspectral spectrum of field samples was reduced to 4 dimensions, the determination coefficient R2 of sheath blight detection model was 0.9685, and the mean square error MSE was 0.0072. Compared with principal component analysis and successive projections algorithm, the results of dimension reduction and detection accuracy of rice sheath blight based on windowed GramSchmidt transform are better, which can provide theoretical basis and technical support for hyperspectral dimension reduction and rice sheath blight control.
作者
曹英丽
肖文
江凯伦
郭宝赢
刘亚帝
王洋
CAO Ying-li;XIAO Wen;JIANG Kai-lun;GUO Bao-ying;LIU Ya-di;WANG Yang(College of Information and Electrical Engineering/Liaoning Agricultural Information Engineering Technology Center,Shenyang Agricultural University,Shenyang 110161,China)
出处
《沈阳农业大学学报》
CAS
CSCD
北大核心
2019年第6期713-721,共9页
Journal of Shenyang Agricultural University
基金
国家重点研发项目(2016YFD0200700,2017YFD0300706)
辽宁省教育厅课题重点项目(LSNZD201605)
关键词
水稻纹枯病
高光谱降维
主成分分析
连续投影法
Gram-Schmidt变换
rice sheath blight
hyperspectral dimensionality reduction
principal component analysis
successive projections algorithm
Gram-Schmidt transform