基于高光谱信息的冬小麦条锈病严重度反演模型通常不能直接应用在宽波段卫星影像上， 而拥有高光谱波段信息的航空遥感影像又因数据尺度小、 成本高难以应用于大规模监测。 文章提出一种通过构建冬小麦条锈病光谱知识库， 利用TM影像实现病情识别和监测的方法。 该方法以包含各种不同病情严重度的试验田的三幅小麦关键生育期PHI航空遥感影像为媒介， 利用病情指数DI的经验反演模型和基于波谱响应函数的TM波段模拟， 建立DI和TM波段模拟反射率间的光谱知识库。 在此基础上， 通过马氏距离法和光谱角度填图(SAM)法将待检象元的光谱信息与光谱知识库进行匹配分析从而实现对病情识别和监测。 监测的精度利用模拟TM象元进行评价， 识别的效果利用TM影像象元进行检验。 结果表明， 该方法在一定生育期范围内具有较佳的监测精度和识别效果。 其中， 使用模拟TM象元在小麦灌浆期精度最佳， 评价的R2达到0.93， 乳熟期次之， 拔节期最差， 基本不能用于监测。 使用TM影像象元在灌浆期和乳熟期可较好地识别染病象元， 在拔节期无法有效识别染病象元。 匹配方法马氏距离法略优于光谱角度匹配法。

In most cases, the reversion model for monitoring the severity degree of stripe rust based on the hyperspectral information can not be directly applied by the satellite images with relatively broad bandwidth, while the airborne hyperspectral images can not be applied for large-scale monitoring either, due to the scale limitation of its data and high cost. For resolving this dilemma, we developed a monitoring method based on PHI images, which relies on the construction of spectral knowledge base of winter wheat stripe rust. Three PHI images corresponding to the winter wheat experimental field that included different severity degree of stripe rust were used as a medium to establish the spectral knowledge base of relationships between disease index (DI) and the simulated reflectance of TM bands by using the empirical reversion model of DI(%) and the relative spectral response (RSR) function of TM-5 sensor. Based on this, we can monitor and identify the winter wheat stripe rust by matching the spectral information of an untested pixel to the spectral knowledge base via Mahalanobis distance or spectral angle mapping (SAM). The precision of monitoring was validated by simulated TM pixels, while the effectiveness of identification was tested by pixels from TM images. The results showed that the method can provide high precision for monitoring and reasonable accuracy for identification in some certain growth stages of winter wheat. Based on the simulated TM pixels, the model performed best in the pustulation period, yielded a coefficient of determination R2=0.93, while the precision of estimates dropped in the milk stage, and performed worst in the jointing stage, which is basically inappropriate for monitoring. Moreover, by using the pixels from TM images, the infected pixels could be identified accurately in pustulation and milk stages, while failed to be identified in jointing stage. For matching algorithms, the Mahalanobis distance method produced a slightly better result than SAM method.