小麦白粉病和条锈病是我国两种最普遍、 最具破坏性的小麦病害, 且田间常常混合发生。 由于病源和发病机理不同, 有必要对这两种病害进行准确区分和识别, 以采取不同的防治措施。 基于ImSpector V10E高光谱成像系统采集的条锈菌侵染叶片、 白粉菌侵染叶片和健康叶片（共计320个）在375～1 017 nm范围内的高光谱图像, 利用高斯平滑等预处理方法得到三种小麦叶片的平均光谱曲线, 发现小麦白粉病和条锈病的敏感波段均集中在550～680 nm的色素强吸收位置, 且趋势基本一致。 针对两种病害的响应波段交叉重叠的问题, 通过主成分分析-载荷法（PCA）、 连续投影算法（SPA）和竞争性自适应重加权算法（CARS）对小麦叶片的光谱信息进行有效降维, 分别优选出3、 6、 30个敏感波段和特征波长; 在此基础上, 采用最小二乘-支持向量机和极限学习机两种分类算法分别基于全波段、 PCA、 SPA和CARS的优选波段, 建立白粉病、 条锈病和健康叶片的判别模型。 结果表明, 8种模型的准确识别率均在94.58%以上。 其中, 主成分分析-载荷法结合极限学习机模型最优, 训练集与校正集的正确识别率分别为99.18%和100%, 且结构简单, 仅含有三个变量（占全波段的1.1%）。 最后, 通过对小麦白粉病、 条锈病以及健康叶片的显微结构分析, 发现病菌入侵叶片, 破环细胞结构, 导致叶绿素含量减少, 光合作用效能降低,进而使得小麦在可见光波段光吸收程度减弱, 反射率增大。 可见, 利用作物的高光谱图像信息能够准确地识别不同类型的小麦病害, 为研发作物病害在线识别的多光谱系统提供重要的理论依据。

Powdery mildew and stripe rust are two of the most prevalent and destructive wheat diseases causing severe decreases in wheat yield in China. It is necessary to quantitatively identify different diseases for spraying specific fungicides. In this study, a line-scanning hyperspectral imaging system (ImSpector V10E) was utilized to capture spectral and imagery information of wheat leaves infected by powdery mildew, stripe rust and normal leaves. Based on 320 hyperspectral images, strong spectral reflectivity responses were discovered at the bands of 550～680 nm in the wheat leaves infected with powdery mildew and stripe rust after the savitzky-golay （SG） smoothing method. To reduce the dimensionality of the spectral matrix, 3, 6 and 30 variables were extracted as sensitive wavelengths from full spectra for different diseases using X-loadings of principal component analysis (PCA), successive projections algorithm (SPA), and competitive adaptive reweighted sampling (CARS), respectively. Least squares support vector machine (LS-SVM) and extreme leaning machine (ELM) were applied to build identification models using full spectra and sensitive wavelengths extracted by X-loadings of PCA,SPA and CARS to distinguish powdery mildew, stripe rust and normal leaves. The accuracy rates of all the models in the calibration set and test set were above 94.58%. Among these models, the ELM classification model combined with X-loadings of PCA had the best performance, with accurate identification rates of 99.18% on the calibration set and 100% on the test set. Moreover, this model was simple in structure with only three variables (560, 680 and 758 nm). Meanwhile, the microstructure of three kinds of wheat leaves were also studied. Although the infection mechanisms of these two diseases were slightly different, they both destroyed the mesophyll cells, reduced chlorophyll content and photosynthesis markedly. The string of changes leaded to weakened light absorption but increased reflectivity in the visible light band. Thus, the results indicated the potential for the rapid and non-destructive detection of wheat diseases by hyperspectral imaging, which could help to develop online multispectral detection system for different kinds of plant diseases.