Evaluation of Stem Rust Disease in Wheat Fields by Drone Hyperspectral Imaging

Abstract

Detecting plant disease severity could help growers and researchers study how the disease impacts cereal crops to make timely decisions. Advanced technology is needed to protect cereals that feed the increasing population using fewer chemicals; this may lead to reduced labor usage and cost in the field. Accurate detection of wheat stem rust, an emerging threat to wheat production, could inform growers to make management decisions and assist plant breeders in making line selections. A hyperspectral camera mounted on an unmanned aerial vehicle (UAV) was utilized in this study to evaluate the severity of wheat stem rust disease in a disease trial containing 960 plots. Quadratic discriminant analysis (QDA) and random forest classifier (RFC), decision tree classification, and support vector machine (SVM) were applied to select the wavelengths and spectral vegetation indices (SVIs). The trial plots were divided into four levels based on ground truth disease severities: class 0 (healthy, severity 0), class 1 (mildly diseased, severity 1-15), class 2 (moderately diseased, severity 16-34), and class 3 (severely diseased, highest severity observed). The RFC method achieved the highest overall classification accuracy (85%). For the spectral vegetation indices (SVIs), the highest classification rate was recorded by RFC, and the accuracy was 76%. The Green NDVI (GNDVI), Photochemical Reflectance Index (PRI), Red-Edge Vegetation Stress Index (RVS1), and Chlorophyll Green (Chl green) were selected from 14 SVIs. In addition, binary classification of mildly diseased vs. non-diseased was also conducted using the classifiers and achieved 88% classification accuracy. This highlighted that hyperspectral imaging was sensitive enough to discriminate between low levels of stem rust disease vs. no disease. The results of this study demonstrated that drone hyperspectral imaging can discriminate stem rust disease levels so that breeders can select disease-resistant varieties more efficiently. The detection of low disease severity capability of drone hyperspectral imaging can help farmers identify early disease outbreaks and enable more timely management of their fields. Based on this study, it is also possible to build a new inexpensive multispectral sensor to diagnose wheat stem rust disease accurately.

Keywords: classification; hyperspectral camera; reflectance; vegetation indices; wavelength.

Figure 1

Top view of the field of the study in Rosemount, Minnesota: (a) aerial image of the entire experiment, blocks 2, 3, and 6 were inoculated, and the other blocks 1, 4, and 5 were sprayed with fungicide; (b) one P. graminis-inoculated block with a superimposed coefficient of infection values.

Figure 2

Unmanned aerial vehicle in the wheat field and calibration panels.

Figure 3

The spectral reflectance curve for wavelength vs. reflectance of wheat plants in stem-rust-inoculated blocks and wheat plants in the fungicide-treated blocks (healthy plants) for (a) binary classification and (b) multiclass (class 1, 2, and 3).

Figure 4

(a) The wavelength ratio or disease sensitivity and (b) the correlation coefficient between healthy and infected plants.

Figure 5

The results of multiclass classification and binary class with error bars among classes using classifiers SVM, DTC, RFC, and QDA. Class 0 is the healthy class, and classes 1–3 are mildly, moderately, and severely diseased classes.

Figure 6

The overall classification accuracy of vegetation indices by applying different classification methods: SVM, DTC, QDA, and RFC.

Figure 7

The variation of fourteen vegetation indices in three different classes.