**1. 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 managemen<sup>t</sup> 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 ofdiseaseseverity(DS),therebyhelpingfarmerstoensuretimelyuseoffungicides[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

**Citation:** Khan, I.H.; Liu, H.; Li, W.; Cao, A.; Wang, X.; Liu, H.; Cheng, T.; Tian, Y.; Zhu, Y.; Cao, W.; et al. Early Detection of Powdery Mildew Disease and Accurate Quantification of Its Severity Using Hyperspectral Images in Wheat. *Remote Sens.* **2021**, *13*, 3612. https://doi.org/10.3390/ rs13183612

Academic Editor: Xanthoula Eirini Pantazi

Received: 12 August 2021 Accepted: 7 September 2021 Published: 10 September 2021

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

frequent mistakes. Laboratory tests include enzyme-linked immunosorbent assay (ELISA), immunofluorescence, fluorescence in situ hybridization (FISH), and polymerase chain reaction (PCR) methods, which are more accurate than visual inspection but usually lag and require precision instruments and strict operations, therefore, most are time consuming.

With the development of spectroscopy technologies, spectral identification methods (Mutka and Bart, 2015) have achieved remarkable results in the field of intelligent identification of plant diseases, such as RGB imagery and multi/hyperspectral technology [6]. A disease identification method using RGB imagery is easier to implement and does not require sophisticated chemical analysis instruments or expensive sensors, imaging equipment, etc. RGB imagery recognition manually extracts the recognizable features of each disease, such as color, shape, and texture features. These features can further be used to obtain a model for distinguishing disease types (Siricharoen) [7]. However, some factors seriously affect the application of visible imaging to plant disease including light condition, removal of the background, and other automatic image processing techniques (Barbedo) [8]. Multispectral images contain more band information compared with RGB images. Satellite imagery and images obtained from the combined use of a UAV with multispectral camera are two common ways to obtain multispectral image. This method seems to be one of the best methods for quickly and accurately detecting and mapping disease incidence over large areas [9]. However, it appears to be insufficient to utilize this combination of narrow bands for disease mapping at earlier stages of the disease's development [10].

Compared with the above methods, hyperspectral technology has a high cost and its complex data interpretation remains challenging [11]. However, its applications are mature and reliable, and have been widely used to monitor the physiological status of vegetation [12], drought stress [13], nutrient deficiencies [14], and disease stress [15]. The principle is that the interaction between plants and the electromagnetic spectrum relies on photosynthetic chlorophyll in the visible light band (VIS), leaf anatomy in the nearinfrared band (NIR), and leaf water in the short-wave infrared band [16]. When wheat is subjected to PM stress, its spectral reflectance changes according to physiological and biochemical changes in its leaves, such as decreased chlorophyll content or destroyed cell structure [17]. When combined with aerial platforms, hyperspectral technology is very well-suited for field phenotyping due to its capability for the characterization of subtle details of diseases. Many studies monitoring plant diseases have been conducted using ground and aerial hyperspectral reflectance. For example, Shi et al. [18] proposed a spectral vegetation indices-based kernel discriminant approach for the detection and classification of yellow rust, aphid, and powdery mildew in winter wheat. Zheng et al. [19] developed optimal spectral indices with three-band combinations based on sensitive wavebands to detect yellow rust disease in wheat at different growth stages. However, by using only spectral information, those studies identified disease at the middle or late infection stages, thereby missing the critical time for prevention and control of the disease.

Hyperspectral imaging (HI) is a new nondestructive monitoring technology developed in recent years. It combines imaging with spectroscopic techniques to obtain both spectral and spatial information [20]. As well as the spectral information of HI, its spatial information offers an accurate and reliable resource to identify disease. Texture is a typical spatial information that refers to the recurring local patterns and permutation rules of pixels. The texture features can reflect changes in the intensity of pixels to distinguish and identify objects [21]. Analyzing texture information from images can detect foliar surface changes due to fungal disease. For example, Yang et al. [22] used gray histogram and a gray-level co-occurrence matrix to extract the texture features of rice blast leaves, establishing a stepwise discriminant model with a classification accuracy of more than 90%. Al-Saddik et al. [23] combined spectral and texture data to detect yellowness and esca in grapevines with an overall accuracy of 99% for both of the diseases. Although the above studies achieved some useful results, there is scant evidence regarding the detection of PM through the fusion of spectral and texture information, especially at the early stages of disease infection.

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.
