基于成像高光谱的小麦冠层白粉病早期监测方法

[目的]本研究利用近地面成像高光谱仪,获取接种白粉病菌后的小麦田间冠层时序影像,探索光谱信息与纹理信息的结合在冠层尺度上早期监测小麦白粉病的能力和表现.[方法]本试验以不同年份、不同抗病性小麦品种的田间试验为基础,利用连续小波(continuous wavelet transform,CWT)方法提取对小麦白粉病敏感的...

Early Detection on Wheat Canopy Powdery Mildew with Hyperspectral Imaging

【Objective】In this study, the near-ground imaging spectrometer was used to obtain time-series images of wheat canopy after inoculation with powdery mildew, which aimed to explore the ability and performance of the combination of spectral feature and texture feature in the early detection of wheat powdery mildew at canopy scale. 【Method】 Based on the field trials of wheat varieties with different disease resistance in different years, the wavelet features sensitive to wheat powdery mildew were extracted by continuous wavelet transform (CWT) method, and the corresponding texture features were extracted based on spectral features to construct normalized difference texture index (NDTI). Meanwhile, the representative traditional vegetation indices (Vis) were selected. Then, based on these features and combinations, the partial least squares discriminant analysis (PLS-LDA) model was used to establish wheat canopy healthy and disease recognition model. The partial least squares regression (PLSR) was used to estimate the severity of wheat canopy disease. The technique was used to distinguish the healthy and disease wheat at different days after inoculation based on the optimal features and combinations. 【Result】 Based on CWT, the selected four wavelet features were 595 nm (yellow region) at 6 scales, 614 nm (red region) at 5 scales, 708 nm (near infrared region) at 3 scales, and 754 nm (near infrared region) at 4 scales respectively. The following texture features were selected for the best texture index combination: ENT₇₅₄, MEA₇₅₄, ENT₇₀₈, ENT₅₉₅, ENT₆₁₄, HOM₇₀₈, HOM₅₉₅, HOM₆₁₄, DIS₅₉₅, HOM₇₅₄ and DIS₆₁₄. Besides, it was found that the texture feature MEA₇₅₄ had the superior performance among all the texture, with the highest correlation between the severity of disease and texture (R²=0.67). The PLS-LDA model based on the combination of wavelet feature and texture feature had the highest accuracy, with the overall classification accuracy of 81.17% and the Kappa coefficient of 0.63. In addition, the PLSR model based on spectral index and texture index was the best, and the R ² of modeling and testing was 0.76 and 0.71, respectively. The severity of wheat canopy powdery mildew was about 26% (about 24 days after inoculation), which was identified in this study at the earliest time. 【Conclusion】 The wheat healthy and disease recognition model based on the combination of wavelet feature and texture feature could significantly improve the accuracy of disease classification, and the combination of spectral index and texture index could significantly improve the accuracy and stability of disease severity estimation. The method and results of this study could provide the reference for disease monitoring of other crops and technical support for accurate application of modern intelligent agriculture.