采用可见/近红外成像光谱技术的玉米籽粒品种识别

Discrimination of corn varieties using visible/near infrared imaging spectrometer system

利用地面成像光谱辐射测量系统(Field Imaging Spectrometer System,FISS)获取了5种玉米籽粒的成像光谱数据,经反射率反演、噪声去除和一阶微分处理后,运用Wilk-lambda逐步判别法进行波段选择并建立判别模型。交叉验证结果表明:(1)平均识别精度为91.6%,除了高油115的识别精度仅有87%外,其他品种的识别精度均在90%以上;(2)在分类方案、光谱波段数、每类样本数量不变的情况下,分类精度受类别数和类别间可分性的影响;(3)随着入选波段数的增加模型识别精度逐步提高。因此成像光谱技术在玉米品种的识别以及质量检测方面具有重要的应用前景。

Hyperspectral images of five corn varieties were acquired using Field Imaging Spectrometer System （FISS）. After reflectance retrieved, noise removal and first-order differential, stepwise discrimination analysis based on the minimization of Wilks＇ lambda was employed to select the feature bands of corn spectral, and then discrimination model was built. The results of Least-one-out Cross- validation （loocv） show that： （1） average discrimination accuracy is 91.6%, in which, discrimination accuracy of High-oil corn No.115 is 87%, and discrimination accuracy of the other varieties is over 90%; （2） if discrimination method, band number and the size of samples of each variety are fixed, discrimination accuracy is effected by variety number and separable; （3） the effect of selected band number on discrimination accuracy is analysed and result shows that discrimination accuracy increases with the increasing of band number. Therefore, FISS has an important application value in com-variety discrimination and quality examination.