高光谱成像的水稻冠层穗颈瘟早期识别

Early Identification of Rice Panicle Blast in Canopy Based on Hyperspectral Imaging

为实现田间水稻冠层穗颈瘟的早期识别,利用室外高光谱成像系统采集早期自然发病大田的水稻冠层穗颈瘟图像,提取、分析反射率光谱特征.对预处理后的高光谱数据,采用主成分分析(Principal Component Analysis,PCA)、植被指数(Vegetation Index,VI)和竞争性自适应重加权法(Competitive Adaptive Reweighted Sampling,CARS)3种方法提取特征变量,结合支持向量机(Support Vector Machine,SVM)和线性判别分析(Linear Discriminant Analysis,LDA)分类算法构建识别模型.结果显示:以CARS特征波长和植被指数构建的模型,从分类结果看都取得了不错的效果,但是特征波长数量较多,可能存在过拟合的风险;单独使用PCA获得的主成分构建水稻冠层识别模型,没有明显效果.为此,研究尝试对选取的植被指数和提取的CARS特征使用PCA进一步降维,得到4个VI-PCs特征和5个CARS-PCs特征用于建模,取得了很好的效果.基于VI-PCs特征的SVM模型和LDA模型的总体分类精度分别为94%和95%;基于CARS-PCs特征的SVM模型和LDA模型总体分类精度分别为95%和97%,实现用较少变量获得较好的区分效果.从模型构建算法来看,LDA算法模型均优于SVM算法模型,说明LDA方法更适合于水稻冠层穗颈瘟识别模型的构建.研究可为航空、航天大面积的作物病虫害遥感监测提供理论依据.

In order to realize the early recognition of rice panicle blast in canopy in the field,the outdoor hyperspectral imaging system was used to collect the images of early stage natural occurring rice canopy panicle blast in the fields,then the spectral characteristics of reflectance were extracted and analyzed.For preprocessed hyperspectral data,PCA,vegetation index and CARS were used to extract feature variables,and build recognition models in combining with LDA and SVM classification algorithms.The results showed that the model constructed by CARS characteristic wavelength and vegetation index achieved good classification results.However,the number of characteristic wavelengths was large,there may be a risk of over-fitting.The principal components obtained by PCA alone were used to construct the rice canopy recognition model,which did not reflect the obvious effect.So,the study tried to use PCA to further reduce the dimension of the selected vegetation index and the extracted CARS features,and obtained 4 VI-PCs features and 5 CARS-PCs features for modeling,which achieved good results.The overall classification accuracy of SVM model and LDA model based on VI-PCs features was 94%and 95%,respectively.The overall classification accuracy of SVM model and LDA model based on CARS-PCs features was 95%and 97%,respectively,which achieved a better discrimination effect with fewer variables.From the view of model construction algorithm,the LDA algorithm model was superior to the SVM algorithm model,showing that LDA method is more suitable for the construction of rice canopy panicle blast identification model.The study can provide a theoretical basis for the aeronautical and astronautical remote sensing monitoring crop diseases and pests over a large area.