应用高光谱成像技术, 结合峰值比判别法和偏最小二乘判别法, 探讨快速无损诊断正常、 缺素和黄龙病柑桔叶片的可行性。 在37428～1 01689 nm可见近红外光谱范围内, 采集了正常、 缺素和黄龙病柑桔叶片的高光谱数据。 以主叶脉为轴线, 两侧各选一个长约60像素、 宽约30像素的椭圆形感兴趣区域。 提取两个感兴趣区域的平均反射率光谱, 经相关分析, 筛选出50279和37428 nm一对特征波长, 建立了正常叶片的峰值比判别模型, 模型误判率为17%, 但该模型无法区分缺素和黄龙病叶片。 采用二阶导数结合平滑光谱预处理方法, 处理反射率光谱, 建立了缺素和黄龙病叶片偏最小二乘判别模型。 采用留一法交互验证确定最佳主成分因子数为17, 建模相关系数为096, 建模标准差为013, 模型对两类叶片分类正确率都达到了100%。 在此基础上, 提出了峰值比判别模型和偏最小二乘判别模型相结合的不同类别叶片二步快速诊断法。 采用未参与建模的正常、 缺素和黄龙病叶片各10片, 评价模型的分类能力, 模型分类正确率达到了967%。 实验结果表明: 应用高光谱成像技术, 结合由峰值比判别模型和偏最小二乘判别模型构成的二步判别法, 快速识别正常、 缺素和黄龙病柑桔叶片是可行的。

The feasibility was investigated for identifying sound, yellow and citrus greening leaves of navel orange trees based on hyperspectral imaging combined with correlation analysis and discriminant partial least square (DPLS) methods. The hyperspectral data of sound, yellow and citrus greening leaves were recorded in the wavelength range of 37428～1 01689 nm. Two regions of interest (ROI) were marked symmetrically on both sides along main veins with an ellipse of major axis of 60 pixels and minor axis of 30 pixels. The average reflectance spectrum was extracted from ROI regions. A pair wavelengths of 50279 and 37428 nm were chosen with correlation analysis method in the wavelength range of 37428～1 01689 nm. The classification model was developed with the peak ratio of the pair wavelengths. This model was effective to sound leaves with the recognition accuracy of 17% but yellow and citrus greening leaves. The DPLS model was employed with the preprocessing spectra of second derivative and Savitzky-Golay smoothing. The recognition accuracy of this model was 100% for citrus greening leaves and yellow ones. The number of latent variables (LVs) was optimized with the leave one out cross validation method. The optimal LVs, correlation coefficient and standard error of calibration of the DPLS model were 17, 096 and 013, respectively. The correction classification rate of the DPLS model was 100% for yellow leaves and citrus greening ones. Two-step method of the peak ratio models combination with the DPLS was proposed for identifying sound, yellow and citrus greening leaves. The new samples were applied to evaluation the classification ability of the two-step method, which included sound leaves of 10, citrus greening leaves of 10 and yellow leaves of 10. The correction classification rate reached 967%. Experimental results showed that it was feasible to identify sound, yellow and citrus greening leaves by hyperspectral imaging coupled with the peak ratio and DPLS models.