**4. Discussion**

*4.1. 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 agreemen<sup>t</sup> 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 grea<sup>t</sup> 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.

**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.

**Figure 11.** Chloroplast structure of wheat at different disease severity (DS) after infection.

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*<sup>2</sup> 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.

### *4.2. 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.

### *4.3. 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.
