小麦条锈病是世界上影响小麦安全生产的一种重要病害。实现小麦条锈病不同严重度叶片快速、准确的分级识别, 对于条锈病监测、预测预报和防治措施的制定具有重要意义。通过人工接种获得条锈病不同发病程度小麦叶片, 选取8个不同严重度级别(1%, 5%, 10%, 20%, 40%, 60%, 80%和100%)叶片各30片和健康小麦叶片30片, 利用近红外光谱技术分别获取光谱信息, 共获得270条近红外光谱曲线, 依据小麦叶片条锈病发病程度的不同, 将其分为9个类别。从每个类别中随机选择7~8条光谱曲线作为测试集, 共计67条, 将剩余的203条光谱曲线作为训练集。利用定性偏最小二乘法建立小麦条锈病不同严重度叶片的定性识别模型。研究分析了不同光谱预处理方法、建模比(训练集:测试集)和建模谱区对所建模型识别效果的影响。结果表明, 在4 000～9 000 cm-1谱区范围内, 原始近红外光谱数据经中心化预处理后, 建模比为3∶1时, 采用内部交叉验证法建模, 训练集和测试集的总体识别准确率分别为95.57%和97.01%, 所建模型识别效果较好。表明基于近红外光谱技术进行小麦条锈病叶片严重度分级识别是可行的, 为小麦条锈病的监测和评估提供了一种新方法。

Wheat stripe rust caused by Puccinia striiformis f. sp. tritici, is an economically important disease in the world. It is of great significance to assess disease severity of wheat stripe rust quickly and accurately for monitoring and controlling the disease. In this study, wheat leaves infected with stripe rust pathogen under different severity levels were acquired through artificial inoculation in artificial climate chamber. Thirty wheat leaves with disease severity equal to 1%, 5%, 10%, 20%, 40%, 60%, 80% or 100% were picked out, respectively, and 30 healthy leaves were chosen as controls. A total of 270 wheat leaves were obtained and then their near infrared spectra were measured using MPA spectrometer. According to disease severity levels, 270 near infrared spectra were divided into 9 categories and each category included 30 spectra. From each category, 7 or 8 spectra were randomly chosen to make up the testing set that included 67 spectra. The remaining spectra were treated as the training set. A qualitative model for identification and classification of disease severity of wheat stripe rust was built using near infrared reflectance spectroscopy (NIRS) technology combined with discriminant partial least squares (DPLS). The effects of different preprocessing methods of obtained spectra, ratios between training sets and testing sets, and spectral ranges on qualitative recognition results of the model were investigated. The optimal model based on DPLS was built using cross verification method in the spectral region of 4 000～9 000 cm-1 when “centralization” was used as the preprocessing method of spectra and the spectra were divided into the training set and the testing set with the ratio equal to 3∶1. Accuracy rate of the training set was 95.57% and accuracy rate of the testing set was 97.01%. The results showed that good recognition performance could be acquired using the model based on DPLS. The results indicated that the method using near infrared reflectance spectroscopy technology proposed in this study is feasible for identification and classification of disease severity of wheat stripe rust. A new method was provided for monitoring and assessment of wheat stripe rust.