**1. Introduction**

The production of wheat plays important social and economic roles, and the quality and safety issues related to these functions have been the focus of research at the national level and abroad. Fusarium head blight (FHB) is a wheat disease caused by the fungus *Gibberella zeae* (*Fusarium graminearum*) and often severely affects wheat yield and quality. Wheat infected with FHB accumulates a large amount of toxins in its grains, thereby seriously threatening public health. These bacterial toxins can contaminate flour and persist in the food chain for long periods, producing carcinogens. Therefore, FHB has become one of the crop diseases of great concern worldwide [1,2].

Conventional crop disease detection methods range from the naked eye to random monitoring, which have the disadvantages of strong subjectivity, high labor intensity, and time consumption. With the rapid development of spectral technology, hyperspectral imaging has been gradually applied

to non-destructive detection of plant diseases and insect pests [3]. Hyperspectral images can provide hundreds of thousands of continuous narrow band data points and are very sensitive to changes in the physical and chemical parameters of plants caused by disease infection. These changes have gradually developed into effective features for expressing plant growth information and have proven to be effective in identifying plant diseases and insect pests [4]. Zheng et al. [5] used wavelengths of 570 nm, 525 nm, 705 nm, 860 nm, 790 nm, and 750 nm to identify yellow rust successfully in the early and middle stages of wheat growth. Huang [6] and others, based on the Relief-F algorithm, proposed that 515 nm, 698 nm, and 738 nm are key wavelengths for distinguishing wheat powdery mildew from other diseases. Bauriegel [7] believe that 550–560 nm and 665–675 nm are the best bands for field identification of FHB. However, with the increase in spectral and image spatial resolution, the simultaneous increase of data dimensions, noise, and redundant spectra pose considerable challenges to data storage, processing, and analysis [8].

The vegetation index is an effective method often used in the field of optical remote sensing to reflect changes in plant physiological and biochemical parameters. This index is a simple and efficient spectral data processing method that combines a few characteristic bands in a certain mathematical form. This method greatly eliminates the redundancy of hyperspectral data, has a small amount of calculation, and is widely used to estimate crop yields [9], pigment content [10], canopy structure [11], and changes in water status [12]. In recent years, exclusive spectral indexes have been proposed and demonstrated unique advantages in plant disease detection. Zhang et al. [8] developed a hyperspectral index based on hyperspectral microscopic images to identify FHB ears with classification accuracy of 0.898. Devadas et al. [13] observed that healthy and susceptible (yellow rust, leaf rust, stem rust) wheat can be distinguished based on the anthocyanin reflectance index. Rumpf et al. [14] combined the spectral index and support vector machine to identify beet leaf spot, leaf rust, and powdery mildew at an early stage, and the classification accuracy was above 0.65. The results of these studies indicate that a spectral index calculated by spectral reflectance at a special wavelength position has high potential for applications in the fields of crop diseases and insect pests. However, these proposed spectral indexes do not clearly indicate the applicable growth stage or only consider a certain growth period of the crop.

The pathological characteristics of wheat after being infected with FHB differ at separate stages, which may cause inconsistent relationships between the spectral index and the status of FHB during different growth periods. In many studies on spectral indexes, researchers have usually pooled observation data at different stages of the entire growth stage to explore characteristic bands and construct spectral indices, thereby weakening the inconsistencies of FHB status in different growth stages [15]. FHB usually occurs during the flowering and filling stages of wheat. At maturity, the damage caused by FHB to wheat yield and quality has been determined. Currently, conducting research on FHB identification is of minimal importance. Therefore, this study focuses on the accuracy and stability of the disease severity monitoring model for FHB at late flowering and early filling stages, with the goal of providing assistance for the scientific control of the disease. The main research objectives are as follows: (1) Based on the difference in spectral responses between healthy and infected ears, the most suitable characteristic wavelengths for identifying FHB were selected and determined by the random forest (RF) technique in the late flowering stage and early filling stage. (2) A Fusarium disease index (FDI) was constructed in the form of the normalized wavelength difference and compared with classical disease index to evaluate the accuracy and stability of FDI.

#### **2. Materials and Methods**

#### *2.1. Wheat Material*

This study was carried out at the experimental base of the Anhui Academy of Agricultural Sciences (31◦89 N, 117◦1 E) in China from 2017 to 2018 (Figure 1). The tested wheat variety was Xinong 979, which is moderately susceptible to FHB. A 10 × 10 m experimental plot was divided into an inoculation area (50 m2) and a control area (50 m2). In the early flowering stage, a small sprayer was used to

spray a freshly prepared spore suspension (*F. graminearum*) on the ears of wheat in the inoculation area. The control area was sprayed with pesticide (Carbendazim, 750 g/hm2) once between the full heading stage and the early flowering stage (18 April 2018) to prevent FHB and ensure a sufficient number of healthy samples for comparative research. Other field management techniques such as fertilization and irrigation were carried out in the two experimental plots according to local agronomic measures. In this study, 149 and 229 wheat ears were collected at the late flowering (3 May 2018) and early filling (9 May 2018) stages for a total of 378 samples.

**Figure 1.** Experimental field plots.

#### *2.2. Inoculum Production*

Under the bench with a sterile environment, the infected wheat grains were treated twice with mercury dichloride–alcohol–sterilized water. The treated grains were added to potato dextrose agar medium, and the culture was grown at 25 ◦C for three days. Five mycelium plugs were picked at the edge of the colony and placed in 100 mL carboxymethyl cellulose medium for four days. The conidia were filtered with two pieces of filter paper and centrifuged at 5000 rpm for 5 min. The concentration of the spore suspension was adjusted to 1 <sup>×</sup> 105/mL with sterile water.

#### *2.3. Data Acquisition and Processing*

#### 2.3.1. Spectral Measurements

The spectral reflectance of the ears was measured using an SOC710E spectrometer (Surface Optics Corporation, San Diego, CA, USA). The spectral range of this instrument is 374–1050 nm, and the spectral resolution is 2.3 nm. After picking wheat ears in the field, the samples were quickly sent to the laboratory in a fresh-keeping box, and the spectral data were collected in a dark room. Wheat ears were placed on a black platform, and the exposure time and the distance between the platform and the lens were adjusted so that the wheat ears could be clearly imaged. Two 75-Watt halogen lamps were placed on both sides of the dark room to illuminate the sample. The hyperspectral imaging system is shown in Figure 2. Measurements on the whiteboard (with a reflectance of approximately one) and dark current (with a reflectance of approximately zero) were performed for spectral correction. The reflectance value of the dark current was recorded by covering the lens with a black cloth. The correction formula is as follows:

$$R = \frac{R\_{\text{original}} - R\_{\text{dark}}}{R\_{\text{white}} - R\_{\text{dark}}} \tag{1}$$

where *R*original represents the original spectral reflectance, *R*dark is the reflectance value of the dark current, *R*white is the reflectance value of the reference whiteboard, and R is the corrected spectral reflectance of the image.

**Figure 2.** Hyperspectral imaging system.
