Wheat Mingxian 169 was grown in plastic pots (7 cm× 7 cm× 8 cm) at a density of 10~15 strains in each pot, in a rust-free growth chamer [(16± 3) ℃, 16 h light/8 h darkness] in the State Key Laboratory of Crop Stress Biology for Arid Areas and College of Plant Protection, Northwest A& F University, China. When the seedling had grown to two or three leaves, the first wheat seedling leaf was gently rubbed with a clean moistened finger to remove the waxy layer from the leaf surface. Puccinia striiformis and Blumeria graminis on naturally infected leaves (from living tissues) were collected and evenly smeared on the first wheat seedling leaf with a brush, and the inoculation zones were labeled with a marker. Each pathogen was inoculated on ten pots, while another ten pots without inoculation were used as healthy controls. The inoculated plants were covered with wet polyethylene bags to maintain 100% relative humidity and were stored for 24 h at 10 ℃ in a dark chamber. Immediately after incubation, plants were transferred to a clean growth chamber set to a diurnal cycle of 16 ℃ for the 16-h light period and 13 ℃ for the 8-h dark period^[14]. On day 15, various scabs were observed on the surface of the inoculated wheat leaves. 80 strip rust leaves, 100 powdery mildew leaves, and 140 healthy leaves were collected.

The hyperspectral images of wheat leaves were captured by a line-scanning hyperspectral imaging systemin reflectance mode. The hyperspectral imaging system was constituted by a visible/infrared imaging spectrometer (ImSpector V10E, Spectral Imaging, Finland), an area array CCD camera with 320× 256-pixel resolution (XEVA2616, XenICs, Belgium), an illumination system with four 150-W quartz tungsten halogen lamps adjusted at an angle of 45° to illuminate the camera’ s field of view, a mobile platform operated by a stepper motor, a camera obscurer, and a computer with data acquisition and preprocessing software (Spectral-Cube data acquisition V10 software). The hyperspectral imaging system was set up in the laboratory with the room temperature of (28± 1) ℃ and relative humidity of 50%. The spectral region was 375~1 017 nm, with a total of 256 bands. To obtain clear and undistorted hyperspectral images, the distance between samples and lens was 65 cm, and the speed of the conveyor was 14 mm· s^-1, and the exposure time was 5 ms during the image acquisition.

Fig.1

	Figure Option ViewDownloadNew Window
	Fig.1 Photographs of wheat seedlings (upper) and schematic diagram of hyperspectral imaging system (lower)

The spectral reflectance of collected samples were collected from hyperspectral images using ENVI+IDL 5.1 software (ITT Visual Information Solutions, Boulder, CO, USA). Rectangles of 5× 4 pixels were selected on inoculation zones of the corresponding leaf as the regions of interest (ROI). The spectra of each pixel in the ROI were averaged, and this spectrum was considered the spectral reflectance of the sample. According to this procedure, a total of 320 mean reflectance spectra of three kinds of wheat samples were obtained and imported into the MATLAB 2016a software (MathWorks, Natick, MA, USA) for further data analysis.

The common processing methods of SG smoothing, multiple scatter correction, standard normal variate transformation, first derivative, and second derivative were tested in this study. SG smoothing was found to eliminate spectral noise while maintaining the spectral characteristics. Therefore, SG smoothing was selected for data treatment throughout the study.

The kennard-stone(KS) algorithm was adopted to divide 80 powdery mildew samples, 100 wheat stripe rust samples and 140 normal wheat samples at a ratio of 3∶ 1 in this study. Afterward, the respective division results were summarized to form calibration sets with 240 samples and test sets with 80 samples.

Full spectra contain all spectral information with large data files, information redundancy, and multiple collinear variables, which could add to the complexity of identification model, reduce the computation speed, and affect the model accuracy. Therefore, it is necessary to reduce the dimensionality of the spectral data to improve the correct identification rate. In this study, PCA, SPA and CARS^[15] were used as optimization methods to select the characteristic wavelengths.

Classification approaches aim to divide the data into a number of distinct classes. LS-SVM and ELM^[16] algorithmwere applied to build classification models to distinguish powdery mildew, stripe rust and normal leaves in this study. And the model performances were evaluated by accuracy rate of identification. Typically, a higher correct identification rate was associated with better model precision. In addition, the number of input variables in the model was used to evaluate whether the model was simple or not, and fewer input variables in the model suggested a simpler model.

The mean spectral reflectivity curves and standard deviation of stripe rust, powdery mildew, and normal wheat samples were shown in Figure 2. It could be seen that the spectra of inoculated and unvaccinated wheat seedlings had similar change trends with wavelength. However, the mean reflectivity of inoculated wheat leaves was higher than that of the healthy samples over the band range of 400~1 000 nm. In particular, there was a strong spectral reflectivity response in the wheat leaves infected with powdery mildew and stripe rust at the band of 550~680 nm, and the greatest reflectance difference could reach 0.15 compared with healthy leaves. In addition, the leaf reflectivity manifested small fluctuations in the band at 750~900 nm, which was largely consistent with the result trend observed by Yuan et al^[17]. It was previously reported that pathogen infection led to structural destruction of chloroplasts, loss of cell water, and weakened the spectral absorption of pigment and water. Thereby, the spectral reflectivity of the infected leaves increased at the visible band and the short-wave infrared band.

Fig.2

	Figure Option ViewDownloadNew Window
	Fig.2 Mean spectral reflectance curves [with standard deviation (SD)] of wheat leaves with powdery mildew, stripe rust and healthy samples covering the range of 375~1 017 nm

2.2.1 Selection of Effective Wavelengths by PCA

PCA was employed to transform the full wavebands (256 wavebands) into several principal components (PCs). Loadings of first PCs were applied for qualitatively identifying the optimal wavelengths that were responsible for the specific features. The first three PCs explained 97% original variations and their score plots were displayed in Figure 3(a). It indicated that three groups of samples provided an apparent clustering and could be distinguished clearly in the score plots. The X-loadings of PC1 to PC3, which revealed the importance of the analyzed variables, were shown in Figure 3(b). The X-loadings values reflected the correlation coefficients of the principal components at different wavelengths. The wavelengths corresponding to the wave peaks or troughs were considered as the characteristic wavelengths. The first three X-loadings plots of PCA indicated that the reflectance at three wavelengths (560, 680 and 758 nm) with the relatively large loading coefficients had the greatest discriminatory effect on healthy and infected wheat.

Fig.3

	Figure Option ViewDownloadNew Window
	Fig.3 Characteristic variables selected by PCA (a): Score projection of the first three principal components; (b): X-loadings plot of the first three principal components on full wavelength region

2.2.2 Selection of Effective Wavelengths by SPA

The number of characteristic wavelengths to be selected was set from 1 to 30, and the optimal characteristic wavelengths were selected based on the root mean square error of calibration (RMSEC). As shown in Figure 4(a), the value of RMSEC decreased rapidly when the number of wavelengths increased from 1 to 5, then decreased slowly with the increase of number of variables. Since too much variables will increase the complexity of model and slow down the computation speed, the number of variables is often selected when the RMSEC value is the minimum or has a relatively small decrease with increasing numbers of variables. In this study, the determined number of wavelengths was 6. The wavelengths selected finally were marked with solid shapes in Figure 4(b): 481, 516, 562, 675, 789 and 838 nm. The selected wavelengths were used as the inputs for the discrimination models.

Fig.4

	Figure Option ViewDownloadNew Window
	Fig.4 Characteristic variables selected by SPA (a): Variation of RMSEC with increasing number of variables in SPA; (b): 6 characteristic wavelengths selected by SPA

2.2.3 Selection of Effective Wavelengths by CARS

The results for variable selection for wheat diseases using the CARS method are shown in Figure 5. Under the action of the exponential decay function, the rate of decrease of variable number became slower with the increase in running times, indicating the stages of roughing and handpicking of the CARS algorithm for key variable selection. The RMSECV value first decreased and then increased, with the smallest RMSECV value at the sampling frequency of 22, which suggested that information variables irrelevant to wheat disease information were eliminated in steps 1~22. As shown in Figure 5, the selected spectral variable subset was optimal at the sampling frequency of 22, which contained 30 spectral variables.

Fig.5

	Figure Option ViewDownloadNew Window
	Fig.5 Characteristic variables selected by CARS

Spectral data of the samples were divided into the calibration set and the test set based on the KS method. Meanwhile, the full spectra, the effective wavelengths selected by PCA, SPA and CARS, were treated as the input variables of the discrimination models. The discrimination models to identify different diseases were constructed by ELM and SVM, and the models were verified using the test set samples. The identification rates of the three kinds of wheat leaves in the calibration set and test set by the proposed model were shown in Table 1. The correct identification rates of the eight models in the calibration set and test set were above 94.58%. It could be seen that the model based on the full spectra (FS) did not display superiority, indicating that the full spectra not only contained useful information but also contained noise. The 30 spectral bands selected by CARS could comprehensively reflect the effective information in the original spectra to distinguish the powdery mildew, stripe rust, and healthy leaves. However, there were only six input variables in the model constructed by SPA and three input variables in the model of X-loadings of PCA, which were only 2.3% and 1.1% of the FS input variable numbers, thus greatly simplifying the recognition model.

Table 1

Table 1

Table 1 Results of identification method for classifying different wheat diseases Classification method Variable selection methods Number of variables Calibration sets detected (undetected) Test sets detected (undetected) Powdery mildew (60) Stripe rust (70) Healthy wheat (110) Accuracy /% Powdery mildew (20) Stripe rust (30) Healthy wheat (30) Accuracy /% SVM FS 256 59 (1) 70 (0) 110 (0) 99.58 20 (0) 30 (0) 30 (0) 100 CARS 30 58 (2) 70 (0) 110 (0) 99.17 19 (1) 30 (0) 30 (0) 98.75 SPA 6 52 (8) 69 (1) 110 (0) 96.25 18 (2) 30 (0) 30 (0) 97.50 PCA 3 47 (13) 70 (0) 110 (0) 94.58 18 (2) 30 (0) 30 (0) 97.50 ELM FS 256 54 (6) 69 (1) 110 (0) 97.92 19 (1) 30 (0) 30 (0) 98.75 CARS 30 56 (4) 70 (0) 110 (0) 98.33 19 (1) 30 (0) 30 (0) 98.75 SPA 6 58 (2) 70 (0) 110 (0) 99.17 20 (0) 30 (0) 30 (0) 100 PCA 3 59 (1) 70 (0) 110 (0) 99.58 20 (0) 30 (0) 30 (0) 100

Table 1 Results of identification method for classifying different wheat diseases

Wheat disease leads to external morphological and internal physiological changes of the leaves. In this paper, the images of cell microstructure of the wheat samples were obtained with a laboratory electron microscope (Olympus-BX53, Olympus Corp., Tokyo, Japan). Figure 6 depicted the microstructure images of mesophyll tissue of normal leaves and leaves affected by stripe rust or powdery mildew. It could be seen that the healthy ones had a complete structure of crosscut sections, a compact structure and plentiful epidermal cells, and the chloroplasts were attached around the cell wall [Figure 6(c)]. The powdery mildew pathogen covered the leaf surface to form the powdery mildew layer, then entered the leaf tissue through a primary germ tube, destroyed the histological structure of the mesophyll cell, deprived the leaf of nutrients, and reduced the water and chlorophyll contents [Figure 6(a)]. Urediniospores landed on the wheat surface, entered the leaf tissue and accumulated to create a large number of summer spore banks or spore beds, after which the histological structure of the mesophyll cell was destroyed and the chlorophyll content reduced [Figure 6(b)].

Fig.6

	Figure Option ViewDownloadNew Window
	Fig.6 Cell microstructure images of wheat leaf tissue infected with powdery mildew (a), stripe rust (b), and healthy wheat as control (c)

Although the infection mechanisms of these two diseases were slightly different, both destroyed the mesophyll cells, reduced chlorophyll content, intensified transpiration, reduced photosynthesis markedly. The string of changes led to weakened light absorption but increased reflectivity in the visible light band. However, the optimal sensitive band for each disease varied according to the disease type, which was consistent with the observations of Mahlein^[2]. Consequently, wheat disease identification could be quantitatively realized through monitoring the characteristic spectral changes of wheat, which could provide timely and reliable monitoring of wheat diseases.