基于近地成像光谱的小麦全蚀病等级监测

小麦全蚀病是检疫性的土传病害,对小麦生产危害极大,对其发生的监测是治理的根本。遥感技术可实时、宏观监测病害发生发展,尤其是成像光谱技术的图谱合一,可精准对病害识别和分类。该文首先通过主成分分析提取不同小麦白穗率的冠层光谱特征;再通过灰色聚类分析方法,研究白穗率等级的可分性;最后利用基于径向基(RBF,radial basis function)核函数的支持向量机对全蚀病害的近地成像高光谱图像进行分类,从而验证近地成像光谱在全蚀病监测上的可行性。研究结果显示:该方法对5种程度的小麦全蚀病白穗率的分类精度均达94%以上,Kappa值大于0.8。研究表明利用该方法,通过近地成像光谱图像可以准确监测小麦全蚀病的病情,对小麦全蚀病的治理有指导意义。

Monitoring and classification of wheat take-all in field based on imaging spectrometer

Abstract: Wheat take-all is a quarantine disease, which will lead to a disaster in wheat production without timely monitoring and management. Remote sensing technique, especially the field-based imaging spectrum technique, can achieve real-time monitoring of the disease development. For rapid extraction of take-all disease information, we try to monitor wheat take-all disease using imaging spectrometer. The experiment was carried out in Baisha village, Yuanyang County of China. We designed test of three concentration gradients and repeated three times, the experimental field was 30 m2. The wheat take-all white head rate was surveyed two weeks before harvest. The wheat's canopy spectrum was collected by two kinds of spectrometer, ASD Handheld non-imaging spectrometer (ASD Handheld, ASD Inc.) and Headwall imaging spectrometer (HyperSpec(r) VNIR, Headwall Photonics, Inc.). All data were collected between 10:00 to 13:00 in sunny days. In this study, based on gray association analysis (GAA) and support vector machine (SVM) classifier, a spectral feature extraction and classification method was proposed to separate the spectral features of the different take-all levels from spectral images. The field-based spectral images were acquired by Headwall imaging sensor. Meanwhile, the spectral data about different white head rate were collected by ASD HandHeld non-imaging sensor. Because of better accuracy and resolution, ASD spectral data had a better capacity to express the spectral features of take-all levels. These spectral features were extracted using kernel principle component analysis (K-PCA). Characteristic bands of the first four of principal component was mainly green band, red band and near infrared band, indicated in the spectrum curve, peak and valley phenomenon was the main distinguishing feature of white head rate and take-all disease grade. Then Jeffries-Matusita distances between feature bands were calculated, if Jeffries-Matusita distances between feature bands were greater than 1.8, the selected characteristic bands can distinguish different damage degree of wheat take-all disease. The spectral separability of take-all levels was tested and assessed by grey association analysis. Based on these significant features, some of Headwall imaging spectral data with different take-all levels were selected as the training data for the field-based spectral images. Through the SVM classifier based on RBF kernel function, a hyperspectral classification image of take-all was calculated. Results showed that the wheat take-all widely existed in the experimental zone, but its distribution had own specific characteristic with different disease levels. The slight disease wheat and the heavy disease wheat were mixture in the experimental zone. The distribution characteristics of serious take-all wheat disease (white head rate greater than 60%) were intensive and block. Slight wheat disease (white head rate between 10%-30%) were widely distributed in the middle of heavy wheat disease(white ear rate between 30%-60%), the proportion of slight wheat disease and heavy heat disease was 29.53% and 26.06%, respectively, very serious wheat take-all disease (white head rate between 60%-90%) and death of wheat disease showed regional distribution in the image, accounted for 10.73% and 19.91%.The overall accuracy of the classification was greater than 94% (Kappa>0.8). To further validate the classification accuracy, field experiment survey data was compared with the spectral classification, misclassification existed mainly in white head rate 30%~40%.These results proves the field-based imaging spectrum has the capacity to achieve the real-time monitoring and classification of wheat take-all condition, and to support the guidance on wheat production.