Early Detection of Powdery Mildew Disease and Accurate Quantiﬁcation of Its Severity Using Hyperspectral Images in Wheat

: Early detection of the crop disease using agricultural remote sensing is crucial as a precau-tion against its spread. However, the traditional method, relying on the disease symptoms, is lagging. Here, an early detection model using machine learning with hyperspectral images is presented. This study ﬁrst extracted the normalized difference texture indices (NDTIs) and vegetation indices (VIs) to enhance the difference between healthy and powdery mildew wheat. Then, a partial least-squares linear discrimination analysis was applied to detect powdery mildew with the combined optimal features (i.e., VIs & NDTIs). Further, a regression model on the partial least-squares regression was developed to estimate disease severity (DS). The results show that the discriminant model with the combined VIs & NDTIs improved the ability for early identiﬁcation of the infected leaves, with an overall accuracy value and Kappa coefﬁcient over 82.35% and 0.56 respectively, and with inconspicuous symptoms which were difﬁcult to identify as symptoms of the disease using the traditional method. Furthermore, the calibrated and validated DS estimation model reached good performance as the coefﬁcient of determination ( R 2 ) was over 0.748 and 0.722, respectively. Therefore, this methodology for detection, as well as the quantiﬁcation model, is promising for early disease detection in crops.


Introduction
Wheat (Triticum aestivum) is an important food crop in China. The sustainable production of wheat is important for the stability of China's society and economy. However, wheat is often attacked by various diseases due to unfavorable environments and management conditions. Powdery mildew (PM) (Blumeria graminis f. sp. tritici) is one of the major diseases that seriously affect wheat yield and quality [1]. Currently, people use disease-resistant varieties [2,3] and spray pesticides [4] to control PM, but these are not optimal solutions, and will produce an increased cost and environmental pollution. The essential approach to combating PM is the early identification and quantitative assessment of disease severity (DS), thereby helping farmers to ensure timely use of fungicides [5].
Traditional methods to identify the disease mainly depend on visual inspection and laboratory tests. Farmers usually visually judge its presence in the field, but there are The main element of HI processing for the detection of disease at its early stages is feature engineering, which transforms the raw hyperspectral data into features suitable for modeling. Feature selection is one of the important methods of feature engineering for dimensionality reduction of hyperspectral data. Feature selection methods, such as stepwise discriminant analysis, select a subset of characteristic wavelengths that preserve certain information of the full spectrum. The selected wavelengths can be used to construct vegetation indices [24] and disease indices [25]. However, most of the feature-selection methods are based on a single dataset or a single model, which do not account for the impaction of the changes of sample or variable on variable selection. To solve these problems, introduced herein is sub-window permutation analysis (SPA) [26]. The main idea of SPA is to build hundreds of sub-models based on different sub-datasets by a random sampling of the total samples and variables, and then to test those models statistically to determine the importance of each variable. SPA is a simple and effective feature-selection method, which considers both the influence of the samples and interaction among the variables. The SPA method was first proposed for chemistry and life science analysis and has proved its advantage. Mo et al. [27] combined near infrared spectroscopy with SPA to select the important variables related to haloxyfop-p-methyl residue in edible oil from the whole band and established a PLS-LDA model with an accuracy of 94.74%. Sun et al. [28] used near infrared (NIR) spectroscopy and SPA to detect complex adulteration of camellia oil. In the field of disease detection, SPA was initiated in wheat PM and successfully extracted information from non-imaging hyperspectral data by Wang [29]. However, the point-by-point spectral characteristics of the entire wheat leaf cannot be analyzed due to the measurement characteristics of the analytical spectral device. In the early stage of wheat infection by powdery mildew, the disease spots are scattered and usually not evenly distributed. For this reason, the combination of spectral and imaging technology can provide a new perspective for the early diagnosis of wheat powdery mildew.
Here, the main objectives of this study were (I) to determine the sensitive features by SPA and combine spectral and texture analysis, (II) to develop a robust discriminant model to early detect wheat PM by PLS-LDA, and (III) to establish an accurate model to quantify disease severity of wheat PM by PLSR. This methodology is expected to provide a new technique to detect the disease. This study demonstrates the capability of hyperspectral imaging to determine occurrence and severity of PM infection at the early stage.

Experimental Site and Design
The experiment was carried out at the Pailou experiment station, Nanjing Agricultural University (32 • 1 N, 118 • 15 E) during the wheat growing season of 2017-2018. Nannong 0686 wheat with high susceptibility to disease was grown in greenhouses. We planted 24 pots (280 mm × 280 mm × 500 mm) with a density of ten seeds per pot. Eight pots were health for control, eight for inoculation at the jointing stage, and the residual eight pots were put separately in the greenhouse for inoculation at booting stage. At the two growth stages, the infected wheat strains were used to shake off spores to inoculate the wheat plant in the pot, then they were placed horizontally in the pot, and all of them were incubated in a controlled environment at a temperature of 25 • C and at 80% relative humidity for 72 h.

Definition of Disease Severity (DS)
The DS of wheat PM is defined as the percentage of the disease pustules portion relative to the total leaf area, which was recorded by visual assessment according to a protocol by Graeff et al. [30]. Because this method is subjective, the assessment was carried out by just one investigator to eliminate any bias due to individual differences. Apart from using DS to quantify the incidence of PM, the leaves are also classified into the categories of healthy leaves (DS less than 1%) and diseased leaves (DS more than 1%) for detection of wheat PM (Figure 1). relative to the total leaf area, which was recorded by visual assessment according t protocol by Graeff et al. [30]. Because this method is subjective, the assessment was carr out by just one investigator to eliminate any bias due to individual differences. Apart fr using DS to quantify the incidence of PM, the leaves are also classified into the categor of healthy leaves (DS less than 1%) and diseased leaves (DS more than 1%) for detect of wheat PM (Figure 1).

Acquisition and Pre-Processing of Hyperspectral Images
In this study, we used the GaiaField portable HI system, consisting of an imag lens coupled with an imaging spectrometer (V10E; Specim, Oulu, Finland) and a C camera (C8484-05; Hamamatsu Photonics, Osaka, Japan) ( Figure 2). The HI system a included two light sources of 150 W halogen lamps (Oriel Instruments, Stratford, C USA) angled at 45°, a computer, a dark box, and a tripod. The spectral data were record in the VIS-NIR range of 400-1000 nm with a spectral resolution of 2.8 nm. The specificat of each hyperspectral image was 1392 × 1040 (spatial dimensions) × 256 (spectral band The distance between the leaves and the camera was set as 300 mm and the exposure ti was set as 0.13 s to obtain clear and non-deformable images. The data were processed a analyzed using the ENVI 5.2 (Exelis Visual Information Solutions, Boulder, CO, USA) a MATLAB R2014a (The MathWorks Inc., Natick, MA, USA) software packages.
The HI preprocessing steps included reflectance conversion, noise removal, a reflectance extraction. Because of the dark current and physical configuration of imaging system, some bands with weaker light intensity contain excessive no Therefore, a raw hyperspectral image was calibrated to reduce the noise according to following equation: where Ic is the calibrated reflectance image, Ir is the raw hyperspectral image, Id is the d current image, and Iw is the white reference image; Id is obtained by covering the cam lens with an opaque cap and turning off the light source, and Iw is acquired by imagin polytetrafluoroethylene white panel with spectral reflectance of 99%. After obtaining the calibrated spectral reflectance, to smooth and minimize the no signals in the images, the noise of the hyperspectral image data was reduced by using minimum noise fraction [31]. The average hyperspectral reflectance of the whole leaf w directly extracted from an image of a single leaf.

Acquisition and Pre-Processing of Hyperspectral Images
In this study, we used the GaiaField portable HI system, consisting of an imaging lens coupled with an imaging spectrometer (V10E; Specim, Oulu, Finland) and a CCD camera (C8484-05; Hamamatsu Photonics, Osaka, Japan) ( Figure 2). The HI system also included two light sources of 150 W halogen lamps (Oriel Instruments, Stratford, CT, USA) angled at 45 • , a computer, a dark box, and a tripod. The spectral data were recorded in the VIS-NIR range of 400-1000 nm with a spectral resolution of 2.8 nm. The specification of each hyperspectral image was 1392 × 1040 (spatial dimensions) × 256 (spectral bands). The distance between the leaves and the camera was set as 300 mm and the exposure time was set as 0.13 s to obtain clear and non-deformable images. The data were processed and analyzed using the ENVI 5.2 (Exelis Visual Information Solutions, Boulder, CO, USA) and MATLAB R2014a (The MathWorks Inc., Natick, MA, USA) software packages.

Selection of the Sensitive Feature
In this study, an SPA algorithm was applied to determine the sensitive wavelength, and then texture features of the corresponding sensitive wavelength were calculated for developing the normalized difference texture index (NDTI).

Construction of Texture Indices
Hyperspectral image has a gray image in each of its spectral bands, so there is a large amount of redundant information. Based on the gray-level co-occurrence matrix (GLCM) of the gray image in each selected spectral band, eight texture features were acquired for system, some bands with weaker light intensity contain excessive noise. Therefore, a raw hyperspectral image was calibrated to reduce the noise according to the following equation: where I c is the calibrated reflectance image, I r is the raw hyperspectral image, I d is the dark current image, and I w is the white reference image; I d is obtained by covering the camera lens with an opaque cap and turning off the light source, and I w is acquired by imaging a polytetrafluoroethylene white panel with spectral reflectance of 99%. After obtaining the calibrated spectral reflectance, to smooth and minimize the noise signals in the images, the noise of the hyperspectral image data was reduced by using the minimum noise fraction [31]. The average hyperspectral reflectance of the whole leaf was directly extracted from an image of a single leaf.

Selection of the Sensitive Feature
In this study, an SPA algorithm was applied to determine the sensitive wavelength, and then texture features of the corresponding sensitive wavelength were calculated for developing the normalized difference texture index (NDTI).

Construction of Texture Indices
Hyperspectral image has a gray image in each of its spectral bands, so there is a large amount of redundant information. Based on the gray-level co-occurrence matrix (GLCM) of the gray image in each selected spectral band, eight texture features were acquired for each gray image [32], as listed in Table 1. In addition to the texture features, a new texture-based index called the normalized difference texture index (NDTI) was proposed, which follows the conventional definition of normalized difference vegetation index. A new NDTI was constructed with all possible two-texture feature combinations at all sensitive wavelengths with eight GLCM-based texture features of each gray image. The texture index NDTI has been used to classify sun and shade leaves of crops and plant biomass estimation [33]. For example, six wavelengths were selected in this research with 8 texture features at each wavelength, thus 48 texture features were acquired. NDTI can be calculated with the combination of two different texture features from all 48 texture features. A total of 2256 combination pairs were obtained, and the specific formula used for NDTI calculation was NDTI = (T1 − T2)/(T1 + T2), where T1 and T2 are the texture features of the selected sensitive wavelength.
Reflects the size of the grayscale change 3 Homogeneity, HOM Reflects local homogeneity of texture 4 Contrast Reflects the clarity of the texture Same as contrast, used to detect similarity 6 Entropy, ENT Measures the amount of information of an image Reflects the uniformity of the grayscale distribution of the image Reflects the extension of a gray value along a certain direction is the value in the cell at row of i and column of j and G is the number of rows or columns. Here, the sensitive wavelengths for each texture feature were selected by SPA. SPA is a new statistics-based feature selection algorithm that combines a random-forest permutation method and model-based cluster analysis. The main aspects of SPA are the establishment of sub-models based on samples and variable sub-windows, and the subjection of parameters of interest to strict statistical analysis. The detailed implementation of SPA can be seen in Li et al. [26].

Selection of Vegetation Indices
When wheat is infected with PM the physiological and biochemical parameters of its leaves change accordingly [16], thereby changing its spectral reflectance. In total, 15 vegetation indices (VIs) frequently investigated in disease research were selected for this study, as listed in Table 2. And these 15 VIs are all highly related to the physiological and biochemical parameters. Table 2. Vegetation indices used in this study.

Development of the Recognition Model for Wheat Leaf Disease
The PLS-LDA method was used to classify the healthy and diseased wheat leaves with the sensitive features of vegetation and texture indices. PLS-LDA is effective in processing data such as small sample size, high dimensionality, and multicollinearity [39]. In the recognition model, the actual measured DS was quantitatively classified into two classes of healthy and diseased for discriminating PM infection. The model performance was evaluated using the overall accuracy (OA) derived from a confusion matrix and Kappa coefficient. An accurate model should have higher values of OA and Kappa coefficient.

Construction of the DS Estimation Model for Wheat Leaf Disease
Furthermore, partial least-squares regression (PLSR) was used to establish the regression relationships between the indices (VIs and NDTIs) and the DS for wheat leaves at different growth stages. As a mature multiple linear regression algorithm, PLSR is the simplest partial least-squares method, and has been widely used to analyze agricultural remote-sensing data [40]. In the severity model, the relationship between actual measured DS and indices was examined to evaluate the sensitivity of indices to PM at different growth stages. The performance of the wheat DS estimation model was evaluated by the coefficient of determination (R 2 ), the root mean square error (RMSE), and the relative root mean square error (RRMSE). R 2 can illustrate the correlation between VIs/NDTIs and DS. The robustness of models was assessed by both RMSE and RRMSE. The higher R 2 and the lower RMSE and RRMSE values, the better the model performs. R 2 , RMSE, and RRMSE can be calculated as following equations: where y i was the observed DS,ŷ i was the estimated DS, y i was the mean of DS observation, and M was the number of samples. The calibration model was built using all data to achieve better performance, but the model may be overfitting. To avoid this phenomenon, 10-fold cross validation was adopted to assess the performance of models. In this process, the data were separated in 10 folds. The 9-fold data were used to build PLSR model. Then, the constructed model was used to predict the remaining 1-fold data. After 10 trials, all DS estimates can be obtained. Figure 3 shows the spectral-reflectance changes of healthy and diseased leaves on different days after infection at the jointing stage. The average DS of the wheat leaves increased gradually over time. Compared with healthy leaves, the spectral reflectance of diseased leaves was higher in the VIS region but lower in the NIR region, which is consistent with previous studies [17]. Moreover, the spectral signature in the NIR region differed more obviously between normal and diseased leaves at late infection stage. consistent with previous studies [17]. Moreover, the spectral signature in the NIR reg differed more obviously between normal and diseased leaves at late infection stage.

Selection of the Sensitive Features
In this study, we used a new feature-selection algorithm SPA for selection of sensit wavelengths for wheat PM, and then we calculated 8 texture features based on the selec sensitive wavelengths for developing the NDTIs. Meanwhile, sensitive VIs were selec and analyzed to compare with the performance of NDTIs. The results of feature select are as the below.

Selection of Sensitive Wavebands
Due to the noisy spectral reflectance data around 400 and 1000 nm, the waveleng used in this study were in the range of 437.2-976.2 nm. By using an SPA algorithm, obtained the conditional synergetic score (COSS) values of each wavelength, indicat its sensitivity to PM, as shown in Figure 4A. The higher the COSS value, the more waveband is sensitive to disease. Then we ranked the spectral bands in descending or according to their COSS values and selected the top 20 bands further for construction PLS-LDA models. We successively added these bands to the PLS-LDA model one by o according to their COSS values in descending order and recorded their calibration a validation accuracies for selection of the best combination of sensitive wavebands. Th calibration and leave-one-out cross validation (LOOCV) accuracies are shown in Fig  4B. Cross validation accuracies gradually increased and achieved highest coinciding po at the sixth waveband, hence, top six wavebands with higher COSS values were selec as sensitive wavelengths, which were 533.5, 659.2, 528.5, 661.7, 968.2, and 523.6 nm descending COSS values. Note: X axis: Wavelength (nm); Y axis: reflectance; "12d (25%)" represents the 12th day after infection with an average DS of 25%.

Selection of the Sensitive Features
In this study, we used a new feature-selection algorithm SPA for selection of sensitive wavelengths for wheat PM, and then we calculated 8 texture features based on the selected sensitive wavelengths for developing the NDTIs. Meanwhile, sensitive VIs were selected and analyzed to compare with the performance of NDTIs. The results of feature selection are as the below.

Selection of Sensitive Wavebands
Due to the noisy spectral reflectance data around 400 and 1000 nm, the wavelengths used in this study were in the range of 437.2-976.2 nm. By using an SPA algorithm, we obtained the conditional synergetic score (COSS) values of each wavelength, indicating its sensitivity to PM, as shown in Figure 4A. The higher the COSS value, the more the waveband is sensitive to disease. Then we ranked the spectral bands in descending order according to their COSS values and selected the top 20 bands further for construction of PLS-LDA models. We successively added these bands to the PLS-LDA model one by one according to their COSS values in descending order and recorded their calibration and validation accuracies for selection of the best combination of sensitive wavebands. These calibration and leave-one-out cross validation (LOOCV) accuracies are shown in Figure 4B. Cross validation accuracies gradually increased and achieved highest coinciding point at the sixth waveband, hence, top six wavebands with higher COSS values were selected as sensitive wavelengths, which were 533. 5

Selection of Optimal Vegetation Indices
The sensitive VIs were selected by correlation analysis between the VIs and DS wheat PM in two growth stages. Figure 5 summarizes the sensitivity of 15 VIs to DS each of the two individual growth stages and at both growth stages as a whole. Except f TVI, these selected VIs were significantly correlated to DS at both stages. The response VIs to DS at different growth stages differed from that of both stages. Most of the V showed high correlation to DS at booting stage, while only a few showed go relationships with DS at the jointing stage. The response of the VIs in different grow stages is not the same. In total, six spectral vegetation indices (PMI, MSR, MCARI, AR SIPI, and PSRI) were significantly correlated with DS at each growth stage as well as f both stages as a whole. They were selected for further disease recognition regression a classification models.

Selection of Optimal Vegetation Indices
The sensitive VIs were selected by correlation analysis between the VIs and DS of wheat PM in two growth stages. Figure 5 summarizes the sensitivity of 15 VIs to DS at each of the two individual growth stages and at both growth stages as a whole. Except for TVI, these selected VIs were significantly correlated to DS at both stages. The response of VIs to DS at different growth stages differed from that of both stages. Most of the VIs showed high correlation to DS at booting stage, while only a few showed good relationships with DS at the jointing stage. The response of the VIs in different growth stages is not the same. In total, six spectral vegetation indices (PMI, MSR, MCARI, ARI, SIPI, and PSRI) were significantly correlated with DS at each growth stage as well as for both stages as a whole. They were selected for further disease recognition regression and classification models.

Selection of Optimal Vegetation Indices
The sensitive VIs were selected by correlation analysis between the VIs and DS of wheat PM in two growth stages. Figure 5 summarizes the sensitivity of 15 VIs to DS at each of the two individual growth stages and at both growth stages as a whole. Except for TVI, these selected VIs were significantly correlated to DS at both stages. The response of VIs to DS at different growth stages differed from that of both stages. Most of the VIs showed high correlation to DS at booting stage, while only a few showed good relationships with DS at the jointing stage. The response of the VIs in different growth stages is not the same. In total, six spectral vegetation indices (PMI, MSR, MCARI, ARI, SIPI, and PSRI) were significantly correlated with DS at each growth stage as well as for both stages as a whole. They were selected for further disease recognition regression and classification models.

Extraction of Texture Features
Because of the redundant information in hyperspectral images, here texture features of images at the six selected sensitive wavelengths were extracted, for the further development of NDTIs. Eight texture features for each spectral band and in total 48 texture features were adopted. Figure 6 shows one of the texture images at the 533.5 nm wavelength. Texture images at same wavelength are various due to different texture features.

Extraction of Texture Features
Because of the redundant information in hyperspectral images, here texture features of images at the six selected sensitive wavelengths were extracted, for the further development of NDTIs. Eight texture features for each spectral band and in total 48 texture features were adopted. Figure 6 shows one of the texture images at the 533.5 nm wavelength. Texture images at same wavelength are various due to different texture features. The Pearson correlation coefficient r was used to examine the relationships between the texture features and the observed DS at different growth stages (Figure 7). The results showed that the texture features were poorly correlated with DS at the jointing stage but highly correlated at the booting stage. Whether it was at the jointing or the booting stage, 20 textures, namely homogeneity (HOM), dissimilarity (DIS), entropy (ENT), and second moment (SEM) at the wavelengths of 523.6, 528.5, 533.5, 659.2, and 661.7 nm were significantly correlated with DS.  The Pearson correlation coefficient r was used to examine the relationships between the texture features and the observed DS at different growth stages (Figure 7). The results showed that the texture features were poorly correlated with DS at the jointing stage but highly correlated at the booting stage. Whether it was at the jointing or the booting stage, 20 textures, namely homogeneity (HOM), dissimilarity (DIS), entropy (ENT), and second moment (SEM) at the wavelengths of 523.6, 528.5, 533.5, 659.2, and 661.7 nm were significantly correlated with DS.

Extraction of Texture Features
Because of the redundant information in hyperspectral images, here texture features of images at the six selected sensitive wavelengths were extracted, for the further development of NDTIs. Eight texture features for each spectral band and in total 48 texture features were adopted. Figure 6 shows one of the texture images at the 533.5 nm wavelength. Texture images at same wavelength are various due to different texture features. The Pearson correlation coefficient r was used to examine the relationships between the texture features and the observed DS at different growth stages (Figure 7). The results showed that the texture features were poorly correlated with DS at the jointing stage but highly correlated at the booting stage. Whether it was at the jointing or the booting stage, 20 textures, namely homogeneity (HOM), dissimilarity (DIS), entropy (ENT), and second moment (SEM) at the wavelengths of 523.6, 528.5, 533.5, 659.2, and 661.7 nm were significantly correlated with DS.

Calculation of Normalized Difference Texture Indices (NDTIs)
NDTIs were constructed to improve the performance of texture analysis in disease recognition as the texture features did not exhibit consistent relationships to DS at two growth stages. Based on the aforementioned 48 texture features, 2256 NDTIs were realized by combining all two possible texture features, and the significance of those NDTIs were analyzed by correlation analysis. Among the 2256 NDTIs, only 10 NDTIs were sensitive to DS at different growth stages and throughout both growth stages, as listed in Figure 8. Ten best performing NDTIs were constructed by a combination of the following seven texture features, namely the second moment (SEM) at 659.2 nm and 661.7 nm and the homogeneity (HOM) at 523.6 nm, 528.5 nm, 533.5 nm, 659.2 nm, and 661.7 nm. Compared with texture features in Figure 7, the NDTIs are better correlated to disease severity. Thus, these NDTIs were selected for further construction of the disease detection model.

Calculation of Normalized Difference Texture Indices (NDTIs)
NDTIs were constructed to improve the performance of texture analysis in disease recognition as the texture features did not exhibit consistent relationships to DS at two growth stages. Based on the aforementioned 48 texture features, 2256 NDTIs were realized by combining all two possible texture features, and the significance of those NDTIs were analyzed by correlation analysis. Among the 2256 NDTIs, only 10 NDTIs were sensitive to DS at different growth stages and throughout both growth stages, as listed in Figure 8. Ten best performing NDTIs were constructed by a combination of the following seven texture features, namely the second moment (SEM) at 659.2 nm and 661.7 nm and the homogeneity (HOM) at 523.6 nm, 528.5 nm, 533.5 nm, 659.2 nm, and 661.7 nm. Compared with texture features in Figure 7, the NDTIs are better correlated to disease severity. Thus, these NDTIs were selected for further construction of the disease detection model.

Evaluation of PLS-LDA Model Based on Different Selected Sensitive Features
The PLS-LDA model was built to evaluate the performance of different indices in terms of detecting wheat PM at different growth stages. The classification results were evaluated both for calibration and validation accuracies, as listed in Table 3. Regardless of growth stage, spectral vegetation indices achieved better classification accuracies than  The PLS-LDA model was built to evaluate the performance of different indices in terms of detecting wheat PM at different growth stages. The classification results were evaluated both for calibration and validation accuracies, as listed in Table 3. Regardless of growth stage, spectral vegetation indices achieved better classification accuracies than texture indices. However, the model based on the combination of VIs and NDTIs performed best in comparison with that of VIs and NDTIs when used alone as input features. The combination of VIs and NDTIs significantly increased the classification accuracy, yielding calibration accuracies of 75.34%, 80.72%, and 76.46% and validation accuracies of 74.88%, 78.93%, and 76.23% at jointing stage, booting stage, and throughout both stages, respectively. The classification accuracies at booting stage were superior, followed by that at both stages and jointing stage, respectively.

Classification of Healthy and Diseased Leaves at Early Stage after Inoculation
As demonstrated in Section 3.3.1 (Table 3), the model based on combination of VIs and NDTIs revealed better results. So, further for classification of healthy and diseased leaves as early as possible after inoculation, we established the PLS-LDA model based on combined features (VIs & NDTIs), on datasets of different days after inoculation (DAI) of each growth stage (Table 4). At the jointing stage, healthy and diseased leaves could be better distinguished on 6 DAI with an average DS of 3.9%, as the overall classification accuracy on this day approached more than 85% with a kappa value of 0.73 and the accuracies for healthy and diseased leaves were 82.35% and 100%, respectively. While, at booting stage, on 3 DAI the overall accuracy reached more than 85% with a kappa value of 0.73 and the accuracies for healthy and diseased leaves were recorded as 81.25% and 100%, respectively. The mean disease severity on this day was noted as 1.1%. Therefore, it can be concluded that wheat PM can be better diagnosed at an early-stage with a maximum of up to 3-6 DAI or at a mean DS of around 1-9%, by using combined (VIs & NDTIs) sensitive features as an input in classification model. Note: DAI indicates days after inoculation, T is mean temperature of green house in degree centigrade, DS is mean disease severity, Healthy and Diseased indicate classification accuracies of inoculated and non-inoculated wheat leaves, OA is overall classification accuracy.

PLSR Model for Estimating the Disease Severity
For estimating the disease severity, the aforementioned six sensitive vegetation indices, 10 texture indices and 16 combined features (i.e., VIs & NDTIs) were used to construct the partial least square regression (PLSR) model with 10-fold cross-validation (Table 5). Across growth stages, the R 2 and RMSE results of booting stage were superior regardless of input features, followed by pooled data of both stages and then jointing stage, respectively. Based on input features, the model constructed by using VIs was comparatively better in the validation model than that of the model constructed by using NDTIs. However, the model constructed by using combined features (i.e., VIs & NDTIs) yielded significantly better results, both for calibration and validation, than the models separately based on VIs or NDTIs.  Figure 9 shows DS estimation models based on combined features (i.e., VIs & NDTIs), by using the validation datasets of different growth stages. It can be observed that the measured DS is close to the predicted DS which lies along line of concordance (1:1 solid line), confirming the accuracy of the DS estimation model. The performance of the estimation model at booting dataset was superior, i.e., the R 2 values between measured and estimated DS was 0.818 and RMSE was 11.320, followed by pooled data of both stages (R 2 = 0.722, RMSE = 13.356). While the model performance at jointing stage was relatively poor (R 2 = 0.532, RMSE = 15.32) and the predicted DS was overestimated. The above results suggest that the PLSR model based on the combined features of VIs and NDTIs performed better for detecting wheat PM disease severity.

Why the Selected Features Are Rational?
In this study, six sensitive spectral bands (533.5, 659.2, 528.5, 661.7, 968.2, and 523.6 nm) were selected by SPA method to reduce the hyperspectral dimensions, in agreement with previous studies revealing sensitive spectral regions to wheat PM [30]. Among these spectral bands, three wavelengths at around 550 nm are sensitive to photosynthetic pigments and have great potential for early disease monitoring, in accordance with the

Why the Selected Features Are Rational?
In this study, six sensitive spectral bands (533.5, 659.2, 528.5, 661.7, 968.2, and 523.6 nm) were selected by SPA method to reduce the hyperspectral dimensions, in agreement with previous studies revealing sensitive spectral regions to wheat PM [30]. Among these spectral bands, three wavelengths at around 550 nm are sensitive to photosynthetic pigments and have great potential for early disease monitoring, in accordance with the findings of Maimaitiyiming [47]. When a pathogen infects wheat, the physiological parameters respond accordingly, such as decreased chlorophyll content, which affects the spectral reflectance in the visible light region. As shown in Figure 10, the chlorophyll content of healthy and diseased leaves on different days after infection were analyzed by analysis of variance (p < 0.05). At the jointing stage, the chlorophyll content of healthy and diseased leaves differed significantly on day 8 after infection. At the booting stage, the chlorophyll content of healthy and diseased leaves differed significantly on day 6 after infection. Decreased chlorophyll content weakened the ability to absorb light, therefore resulting in higher reflectance in the VIS range.

Why the Selected Features Are Rational?
In this study, six sensitive spectral bands (533.5, 659.2, 528.5, 661.7, 968.2, and 52 nm) were selected by SPA method to reduce the hyperspectral dimensions, in agreeme with previous studies revealing sensitive spectral regions to wheat PM [30]. Among the spectral bands, three wavelengths at around 550 nm are sensitive to photosynthe pigments and have great potential for early disease monitoring, in accordance with t findings of Maimaitiyiming [47]. When a pathogen infects wheat, the physiologi parameters respond accordingly, such as decreased chlorophyll content, which affects t spectral reflectance in the visible light region. As shown in Figure 10, the chloroph content of healthy and diseased leaves on different days after infection were analyzed analysis of variance (p < 0.05). At the jointing stage, the chlorophyll content of healthy a diseased leaves differed significantly on day 8 after infection. At the booting stage, t chlorophyll content of healthy and diseased leaves differed significantly on day 6 af infection. Decreased chlorophyll content weakened the ability to absorb light, therefo resulting in higher reflectance in the VIS range. In the near-infrared range, the reflectance is affected by leaf anatomy, whi corresponds to the sensitive wavelength at 968.2nm in this study. As shown in Figure the chloroplast structure (marked by "Ch") in leaves with different DS on 13 DAI w observed using a transmission electron microscope. As the DS increased, the chloropl in the wheat leaves gradually swelled in the cells. Cavities of varying sizes appear surrounded by a single layer of membrane, and the chloroplast disintegrated eventual Figure 10. Changes and comparison of chlorophyll content between healthy leaves and diseased leaves infected at jointing stage (A) and booting stage (B) on the day after inoculation (DAI). Note: ab represents a difference that is significant (p < 0.05).
In the near-infrared range, the reflectance is affected by leaf anatomy, which corresponds to the sensitive wavelength at 968.2nm in this study. As shown in Figure 11, the chloroplast structure (marked by "Ch") in leaves with different DS on 13 DAI was observed using a transmission electron microscope. As the DS increased, the chloroplast in the wheat leaves gradually swelled in the cells. Cavities of varying sizes appeared surrounded by a single layer of membrane, and the chloroplast disintegrated eventually. The cells in the normal leaves form multiple scattering of light resulting in high reflectance, while powdery mildew affects the internal structure of leaves, thus leading to lower reflectance of infected leaves. The cells in the normal leaves form multiple scattering of light resulting in high reflectance, while powdery mildew affects the internal structure of leaves, thus leading to lower reflectance of infected leaves. For texture features, Xie et al. [48] first compared spectral and texture features for detecting different diseases on tomato leaves, and the results obtained by texture features (60.2%) were slightly worse than those obtained by reflectance (97.1%) in classification accuracy. Al-Saddik et al. [23] used hyperspectral images to detect yellowness and esca in grapevines by combining spectral and texture data with an overall accuracy of 99% for both of the diseases. The adoption of NDTI in disease detection of this study is inspired by Zheng [33], who combined NDTIs and VIs to estimate above ground crop biomass with R 2 of 0.78. However, the features selected in this study are specific to the wheat disease PM. When this methodology is used to monitor diseases on other crops, it would need to re-select the sensitive spectral features and corresponding texture features, as the symptoms (shape and color) of each disease are different.

What Is the Reason of Detection Performance at Varied Growth Stages?
The detection at the booting stage performed significantly better than that of jointing stage, which might be due to the different rates of disease propagation at two growth stages. Most of the DS values for diseased leaves obtained after infection at the jointing For texture features, Xie et al. [48] first compared spectral and texture features for detecting different diseases on tomato leaves, and the results obtained by texture features (60.2%) were slightly worse than those obtained by reflectance (97.1%) in classification accuracy. Al-Saddik et al. [23] used hyperspectral images to detect yellowness and esca in grapevines by combining spectral and texture data with an overall accuracy of 99% for both of the diseases. The adoption of NDTI in disease detection of this study is inspired by Zheng [33], who combined NDTIs and VIs to estimate above ground crop biomass with R 2 of 0.78. However, the features selected in this study are specific to the wheat disease PM. When this methodology is used to monitor diseases on other crops, it would need to re-select the sensitive spectral features and corresponding texture features, as the symptoms (shape and color) of each disease are different.

What Is the Reason of Detection Performance at Varied Growth Stages?
The detection at the booting stage performed significantly better than that of jointing stage, which might be due to the different rates of disease propagation at two growth stages. Most of the DS values for diseased leaves obtained after infection at the jointing and booting stages were 1-30% and 1-50%, respectively. The factors which affect the development of wheat PM are environment conditions [49], variety resistance [3], and the number of pathogen spores [50]. The varieties, cultivation practices and the number of spores for inoculations used in the two growth stages were the same. Hence, the difference was mainly due to the different climatic conditions during the two inoculation periods. The inoculation interval between the two growth stages differed by 14 days. The inoculation time at the jointing stage was early spring, at this stage the temperature was still low and gradually rising, therefore the development of the disease was slow. At low temperature, the PM fungus tends to reproduce in small quantities and the physiological and biochemical characteristics of diseased plants may not be obvious. However, the disease developed rapidly at the booting stage because of the suitable environment conditions. It was the time of late spring and conditions of high temperature and humidity were optimal for disease development, consistent with the observations of Yao et al. [51]. Such disease favorable environmental conditions enhanced the rate of disease development at the booting stage, thus inducing obvious changes in the biophysical and biochemical parameters of plants which resulted in changed spectral responses [52]. Moreover, high lesion ratios on the leaf surfaces at booting stage changed the texture patterns of wheat leaves more severely thus, ensuring robust and accurate disease detection at booting stage.

How Early to Detect the Disease When the Combined Feature Is Applied?
This paper focuses on the methodology with regard to the feasibility of detecting the disease at its early stage or even before symptoms appear, at which it is difficult to acquire the weak signal due to the limited bands and course resolution. To the best of our knowledge, the present work is the first attempt to detect wheat PM by combining spectral vegetation indices and texture indices. Wheat PM lesions induce modification of color and surface properties of wheat leaves, and the texture information describes the intensity of the pixel's changes in an image. This phenomenon explains the robustness of the texture approach combined with the spectral approach in the better diagnosis of disease. Previous studies found that the spectral approach performed better when the DS is usually greater than 20%, with classification accuracy of about 87%. Meanwhile, the texture approach performed better in distinguishing DS of less than 20% with classification accuracy of 84%, but the accuracy by using a combination of texture and spectral features can reach up to 90% [53]. Therefore, combining spectral and texture information allows significantly better classification accuracy of diseased and healthy leaves and DS estimation at the early stage of wheat PM.
The early detection of disease on a daily basis after inoculation at different growth stages was carried out by combining the spectral and texture indices, and the classification accuracies of healthy and diseased leaves increased with increasing DS. On 6 DAI at jointing stage, the classification of healthy and diseased leaves was robust and stable, when the average DS was 3.9%. While at the booting stage, disease was recognized well on 3 DAI with an average DS of 1.1%. As the DS increased, the spectral and texture features changed, and the classification accuracies improved as the days passed after inoculation. However, at 9 DAI of booting stage, a slight decrease in classification accuracy was observed. These slight variations in classification accuracies were due to the unbalanced number of healthy and diseased leaves. At the booting stage on 9th DAI, most of the leaves were infected, so the model tended to predict the diseased one, resulting in decreased accuracy of healthy leaves. In another aspect, the variations of chlorophyll contents in leaves are also consistent with the model's performance for disease detection, as the model correctly identified disease on 6 and 3 DAI (Table 4) with an OA of over 85% and chlorophyll contents significantly changed between healthy and diseased leaves on 8 and 6 DAI (Figure 10), at jointing and booting stages, respectively. Therefore, our results well support the idea that HI is an efficient and reliable method for plant disease detection at early stage. It is worth mentioning that ground-based hyperspectral analysis is not suitable for a large area, and it is too expensive to be used by average farmers compared with UAV platforms. Multispectral UAV platforms can monitor disease in large areas with stable and high efficiency but perform poorly at the early disease infection stage. In the future, we believe there would be more hyperspectral equipment mounted on UAV systems, however currently there is only one kind of UHD185 in the community.

Conclusions
In this paper, spectral and texture analysis of hyperspectral imaging technique were used to discriminate wheat powdery mildew without obvious symptoms and estimate corresponding disease severity. This study indicates that the combination of spectral and texture approaches is the sensitive feature for disease detection when building a PLS-LDA classification model and PLSR DS estimation model, which offered significantly improved accuracies in detection and quantification of wheat PM disease. The model based on selected sensitive features identified the disease even before the initiation of significant variations in the physiological and biochemical parameters of leaves after disease. Therefore, it can provide the basis for early detection, preventing and controlling the plant disease worldwide. The stable performance of the spectral and texture indices enabled the early detection of disease with more than 85% overall classification accuracies of PLS-LDA model, especially at early infection stage of 3-6 DAI with a DS of 1-6%. The DS can be well estimated by PLSR model with the R 2 value of 0.818 at booting growth stage. However, the present study is conducted on leaf scale with relatively few trials, and only used limited number of diseased and healthy leaves for classification on a daily basis after inoculation. In the future, we would evaluate the feature and model with a greater number of trials and data, and hope to expand to field and canopy level.