*3.1. Spectral Response Di*ff*erences of Wheat FHB with Di*ff*erent Disease Severities*

In the process of using hyperspectral data to accurately identify diseases, the study of spectral signatures under different disease severities is the basis for screening and identifying sensitive bands of diseases. Figure 4 shows the spectral signature of wheat ears with different infection levels. Generally, in the range of the 550–720 nm band, the spectral reflectance of healthy ears is lower than that of infected samples, with an obvious green peak and red valley; accordingly, these two spectral features disappear in the severely infected ears. Conversely, in the range of the 721–1000 nm band, the more severely infected the sample, the lower its reflectivity. The difference in the responses of wheat ears with different severities in the 550–720 nm and 721–1000 nm bands may be related to the difference in the pigment content and moisture content in mesophyll tissue [9]. Furthermore, with the increase in the severity of wheat diseases, a clear red edge moved in the short-wave direction. The above obvious spectral signature differences provide an important optical basis for analyzing and constructing the relationship between the spectral index and FHB severity in this study.

**Figure 4.** Spectral reflectance curves of wheat ears with different disease severities.

## *3.2. Construction of Proposed New Spectral Disease Index for Identifying Wheat FHB*

### 3.2.1. Characteristic Bands for Identifying FHB at Different Growth Stages

RF was used to select characteristic wavelengths in samples during the late flowering stage, early filling stage, and combination of both. The weight coefficients of all wavelengths were calculated in the spectral range of 374–1050 nm. Except for the extreme points, the weight coefficients of adjacent wavelengths were similar, thus indicating that the information of adjacent wavelengths was highly correlated. To reduce the redundant information and maximize the effective spectral information, this study selected the wavelength corresponding to the positive highest weight coefficient and the negative lowest weight coefficient as the characteristic wavelengths. As shown in Figure 5, the characteristic wavelengths were 570 nm and 678 nm at the late flowering stage, 565 nm and 661 nm at the early filling stage, and 560 nm and 663 nm at the combined stage.

**Figure 5.** Weight coefficients calculated by RF at the late flowering stage (**a**), early filling stage (**b**) and combined stage (**c**).

Characteristic wavelengths selected from the different stages were all in the range of 565–680 nm. Combined with Figure 5, this band range shows a significant difference between healthy and infected wheat ears. Furthermore, the characteristic wavelengths selected in the two growth stages were different, from 678 nm and 570 nm in the late flowering stage to 661 nm and 565 nm in the early filling stage. With the development of FHB, the position of the characteristic wavelength moved in the direction of the short wave, which was consistent with the spectral change of the previously reported disease stress plants; specifically, the red edge shifted in the direction of the blue wave. Notably, the characteristic wavelengths of the combined stage are closer to those of the early filling period, which may be due to the sample size in the early filling stage (229) being larger than that in the late flowering stage (149); moreover, the incidence characteristics of FHB were obvious in this growth period.

#### 3.2.2. Construction of New Fusarium Disease Index for Identifying Wheat FHB

In this study, with FDI as the independent variable and SI as the dependent variable, the relationship between FDI and SI in different stages was evaluated by linear regression analysis (Figure 6). FDI made an accurate prediction of the SI of wheat ears at the late flowering stage, early filling stage, and combined stage (*R*<sup>2</sup> was greater than 0.90, RMSE was less than 0.08). At each stage, the *R*<sup>2</sup> and RMSE of the training and test datasets were close, indicating that the model had a strong generalization ability. From the results of the training and test datasets, the FDI prediction was the most accurate in the early filling stage, followed by the late flowering period, and the lowest in the combined stage. (0.96, 0.94, and 0.90, respectively).

In this study, the regression model obtained from the combined stage was applied to the test set of the late flowering and early filling stages (Figure 7). The results obtained by applying the regression model established through the combined stage to the test datasets of the late flowering and early filling stages (*R*<sup>2</sup> = 0.91 and 0.94, respectively) were slightly lower than those of the regression models of the late flowering and early filling stages (*R*<sup>2</sup> = 0.94 and 0.96, respectively), especially in the late flowering stage.

**Figure 6.** Evaluation of regression models in the training and test datasets at the late flowering stage (**a**,**b**), the early filling stage (**c**,**d**) and the combined stage (**e**,**f**).

**Figure 7.** Evaluation of regression models at the combined stage used at the late flowering stage (**a**) and early filling stage (**b**).

#### *3.3. Comparison of FDI and Traditional Spectral Indices*

To verify the application potential of the FDI for detection of FHB, the results were compared with 16 other published spectral indexes at different stages (Tables 2–4). In the late flowering stage and combined stage, only the FDI proposed in this study had an *R*<sup>2</sup> above 0.9 in the training and the test datasets. In the early filling stage, only FDI and NRI had an *R*<sup>2</sup> above 0.9 in both the training and test datasets. The characteristic wavelengths of FDI (661 nm, 565 nm) and NRI (670 nm, 570 nm) were also close. The prediction results of FDI were higher than those of other spectral indexes in different stages, especially in the late flowering stage, which indicated that the FDI had excellent monitoring accuracy in the early

stage of FHB infection. Furthermore, the detection capabilities of the selected spectral indices were different at separate stages, but the *R*<sup>2</sup> of FDI at every stage was greater than 0.9. Among these indices, nitrogen reflectance index (NRI), transformed vegetation index (TVI), and green index (GI) performed relatively well at different growth stages. What they all have in common is that they have better predictions during a single growth period than during the combined growth period. The diversity of samples in the combined stage may increase the difficulty of model prediction in this period because the characteristic wavelengths are dynamically changing in different growth stages [5]. At the same time, the performance of some indices in different growth stages will be very different. For example, modified chlorophyll absorption in the reflectance index (MCARI) has much better predictive power in the early filling period (*R*<sup>2</sup> = 0.67) than in the late flowering period (*R*<sup>2</sup> = 0.41); normalized pigment chlorophyll ratio index (NPCI) has much better predictive power in the combined period (*R*<sup>2</sup> = 0.77) than in the early filling period (*R*<sup>2</sup> = 0.17). The same index responds differently to diseases in different growth stages, which may be affected by the pathogenic mechanism of vegetation [13].






**Table 3.** *Cont*.

**Table 4.** Comparison of FDI and traditional spectral indices at the combined stage.


#### **4. Discussion**

#### *4.1. Analysis of Spectral Characteristics for Identifying Wheat FHB*

Previous studies have demonstrated that changes in crop physiological and biochemical parameters lead to changes in spectral reflectance, which is the basis for optical technology used to diagnose the severity of FHB [5]. As the symptoms of FHB are different in separate growth stages of wheat, characteristic wavelengths used to identify the severity of FHB are also different, so it is necessary to extract these wavelengths for different stages. In addition, there are some differences in the spectra of wheat ears with different degrees of infection in specific bands. Figure 4 shows the destruction of chloroplasts in the ear tissue of FHB which causes the chlorophyll in the cells to continuously degrade, and the reflectivity of the spectrum in the chlorophyll band (560–675 nm and 682–733 nm) decreases rapidly. At the same time, the decrease of the chlorophyll content in these cells reduces the possibility of photon reemission and reabsorption in this wavelength range, resulting in an increase in spectral reflectivity and the blue shift of the "red edge". These changes become more obvious with the increase in the degree of infection. The characteristic wavelengths selected in this study are in the range of 560–680 nm, including the green reflection peak and red absorption valley, and can characterize the characteristics of wheat ears. According to the characteristic wavelengths of the late flowering stage (570 nm and 678 nm) and early filling stage (565 nm and 661 nm), the 570 nm

and 565 nm wavelengths are near the green peak, while 678 nm and 661 nm are near the red valley. Therefore, the characteristic wavelengths selected in this study are key wavelengths for identifying wheat ears with FHB. Bauriegel [7] also confirmed the importance of this band in the early detection of wheat ear scabs.

#### *4.2. Comparison of Application E*ff*ects between Proposed New FDI and Traditional Spectral Indices*

The most common and serious symptom of wheat FHB, ear rot, often begins at the early flowering stage. Therefore, the monitoring of wheat FHB at the flowering stage is highly valuable. However, the FHB fungi in the flowering stage are in the stage of mass reproduction, and the physiological and biochemical characteristics of the infected ears are not obvious, which makes it more difficult for a conventional spectral index to detect the severity of FHB at the flowering stage. The 16 published spectral indices selected in this study were constructed by collecting sample data from different stages of multiple growth stages, so their accuracies in detecting the severity of FHB at the late flowering stage were relatively poor. In this study, the sample models of the late flowering stage and combined stage were applied to the test set of the late flowering stage. The results show that the model obtained from the late flowering stage sample was more suitable for the detection of FHB at the late flowering stage than the model obtained from the combined stage sample. The characteristic wavelength selected from the samples at the late flowering stage (570 nm and 678 nm) was therefore more suitable for FHB detection than the characteristic wavelength selected during the combined stage (560 nm and 663 nm).

In the sample verification of the combined stage, the accuracies of the 16 published spectral indexes were not satisfactory because none of them were able to achieve an *R*<sup>2</sup> above 0.85 in both the training and test datasets. These spectral indices are mainly based on the leaves or canopies of crops rich in chlorophyll, and few diagnostic studies have used these indices to evaluate the severity of FHB in wheat ears, which is a special part with a low chlorophyll content. In the early filling period, in addition to the FDI, the NRI, plant senescence reflectance index (PSRI), GI, normalized difference vegetation (NDVI), and optimized soil-adjusted vegetation index (OSAVI) also had accurate detection results. This may be because the morphology and cell structure of wheat ears caused by FHB are more obvious than those at the late flowering, so other spectral indices may be more accurate in this context. Notably, among these spectral indexes, NRI and GI both performed strongly at the late flowering and early filling stages but performed relatively poorly in the combined stage. This shows that the diversity of samples has an important effect on predictions of FHB severity, which demonstrates the importance of subdividing the growth period when exploring the forecasts of FHB severity.

#### *4.3. Analysis of Other Influential Factors*

Hyperspectral data contain hundreds of narrow-band data points, but the adjacent wavelength information is often highly correlated, so the use of full-band information will only increase the complexity of data acquisition and calculation [5]. Usually, the most effective information is only contained in some specific bands, and the rest is redundant information [36]. In addition, the high price of hyperspectral imaging systems will also limit the application potential of the technology. This study explored the characteristic wavelengths at different stages to design a multispectral camera with a low price, a fast processing speed, and wide applications for specific identifications of FHB in different growth stages. In this study, the *R*<sup>2</sup> of predicted SI and FDI exceeded 0.90 in the late flowering, early filling, and combined stages. Considering the influence of man-made or natural environmental factors, this prediction accuracy is acceptable.

In addition, the spatial distribution and severity of FHB diseases are greatly affected by the genetic resistance of different varieties as well as environmental and agricultural management factors. Therefore, the factors that affect the biophysical and biochemical parameters of plants will affect the identification of FHB. This study was conducted under laboratory conditions, and its direct application in the field requires further verification. In the future, the growth stage will be further subdivided,

especially in the early stages of FHB occurrence, such as mid-flowering, late flowering, early filling, and mid-filling, to achieve early detection, protection, and evaluation.

#### **5. Conclusions**

Monitoring wheat infection by FHB at different growth stages is important in making a decision on the use of pesticides to protect wheat from FHB and to evaluate yield losses. In this study, RF was used to select characteristic wavelengths for the late flowering stage, early filling stage, and combined stage of both. These wavebands were 570 nm and 678 nm for the late flowering stage, 565 nm and 661nm for the early filling stage, 560 nm and 663 nm for the combined stage. In the light of above wavelengths, FDI at each stage were constructed for establishing linear regression models with SI. Every model showed a high predictive accuracy with the test datasets, with their *R<sup>2</sup>* values exceeding 0.90. In addition, the *R<sup>2</sup>* of the model established at the late flowering stage and early filling stage was better than that of the combined stage, and the *R<sup>2</sup>* of applying the model of the combined stage to the test dataset at the late flowering stage and filling stage also decreased. Therefore, it is indicated that FHB shows different spectral characteristics at each growth stage, which provides a favorable basis for detecting the severity of the FHB disease at different growth stages in the future. However, additional studies are needed to verify the universality of FDI on different wheat varieties and in different field experiment settings.

**Author Contributions:** All authors contributed to the study conception and design. Methodology: F.L.; Writing—review and editing: H.Q.; Formal analysis and investigation, Writing—original draft preparation: D.Z.; Formal analysis and investigation, Writing—original draft preparation: Q.W.; Software: X.Y.; Conceptualization: C.G. All authors have read and agree to the published version of the manuscript.

**Funding:** This research was funded by the National Natural Science Foundation of China (Grant No. 41771463 and 41771469), Anhui Provincial Major Science and Technology Projects (Grant No. 18030701209), and National Key Research and Development Program of China (Grant No. 2016YFD0300700).

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2020 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 (http://creativecommons.org/licenses/by/4.0/).
