Wheat is the second largest grain crop in China. The stable and high yield of wheat has been the focus of agricultural production (Huang et al., 2020). Fusarium head blight (FHB), also known as scab, is a devastating wheat disease caused by the fungal plant pathogen Fusarium graminearum (Gibberella). Fusarium-infected wheat typically results in small, low mass, and shrunken grains, which can rapidly lead to very large crop losses and quality degradation (Bauriegel et al., 2011b). Furthermore, the fungus produces a large number of mycotoxins (deoxynivalenol, nivalenol and zearalenones etc.), among which the most toxic deoxynivalenol (DON) can disrupt normal cell function by inhibiting protein synthesis, posing a significant threat to human and animal health (Barbedo et al., 2015). In recent years, with global climate change, wheat FHB infection has become increasingly serious, resulting in severe damage to wheat quality and yields. Ineffective FHB management practices hinder the profitable and sustainable production of wheat, affecting its economic and social benefits in China. Therefore, the detection of disease development of wheat is important and essential for successful disease control.

Traditional FHB detection mainly relies on professionals to scout the development of wheat infection through visual interpretation, or scholars use chemical methods, such as gas chromatography (GC) (Simsek et al., 2012), high performance liquid chromatography (HPLC) (Simsek et al., 2012), enzyme-linked immunosorbent assay (ELISA) (Maragos et al., 2006), and polymerase chain reaction (PCR) (Amar et al., 2012; Atoui et al., 2012) to detect FHB and DON production. However, these methods are time-consuming, labor-intensive, unable to achieve large-scale monitoring, and are destructive to wheat. Remote sensing technology has been widely used in the monitoring and identification of wheat FHB with nondestructive inspections and rapid measurements. At present, monitoring of wheat FHB using remote sensing technology is mainly manifested in three aspects: (i) identify wheat kernels with varying degrees of damage under laboratory conditions to accurately judge the quality of wheat kernels (Delwiche et al., 2011; Barbedo et al., 2015; Jaillais et al., 2015; Alisaac et al., 2019; Femenias et al., 2020; Liang et al., 2020; Zhang D. Y. et al., 2020a; Zhang D. Y. et al., 2020b); (ii) use remote sensing technology to capture the information of individual or canopy wheat infected with FHB to accurately detect the disease (Dammer et al., 2011; Menesatti et al., 2013; Whetton et al., 2018a,b; Huang et al., 2019b; Zhang et al., 2019; Huang et al., 2020; Ma et al., 2020; Huang et al., 2021); and (iii) monitor wheat FHB on a regional scale with remote sensing (Liu et al., 2020b). However, there are many limitations in these studies. The inspection of wheat kernels has a time lag that only allows the use of kernels with different qualities and cannot fundamentally ameliorate wheat infection. Quantitative detection studies at the single plant scale or canopy scale only provides a theoretical reference without the spatial distribution of wheat infection to meet the needs of practical applications. Optical satellite images are at risk of being covered by clouds, and FHB may occur severely and frequently in cloudy and foggy areas, reducing the availability of remote sensing images (Liu et al., 2020a). Therefore, there is an urgent need for new technological means to solve the current problems.

Unmanned aerial vehicles (UAVs) are considered a practical detection method for crop pests and diseases. Unlike near-ground and satellite-based remote sensing platforms, applications of UAV have the advantages of large coverage, high efficiency, and flexibility (Fu et al., 2022; Zhu et al., 2022a). UAV can collect very high-resolution images and data in a cost-effective manner over a short period of time (Ye et al., 2020). As a new technological means, UAV technology has made significant progress in crop classification, growth monitoring, and identification of pests and diseases. UAV also allows for a proper balance between image quality, sensing efficiency, and operating cost (Li et al., 2019). At present, UAV images are mainly divided into multispectral images and hyperspectral images. Hyperspectral images have dozens to hundreds of continuous and subdivided spectral bands in the ultraviolet, visible, near-infrared, and mid-infrared regions, making them more sensitive to the reflected energy of light and increasingly available (Liu et al., 2020a). Hyperspectral images can provide image and spectral data of each pixel, thus detecting the internal chemical compositions and external phenotypic traits of objects (Zhang D. Y. et al., 2020b). Currently, there are few reports on the detection of FHB infection in wheat using UAV hyperspectral technology (Liu et al., 2020a; Ma et al., 2021; Xiao et al., 2021). We attempted to use UAV hyperspectral technology to explore wheat FHB detection methods in our study. What’s more, scholars have primarily mined spectral features that could characterize physiological and biochemical changes (such as moisture, pigment, etc.), as well as considered texture features that can represent spatial changes of wheat to detect FHB. In fact, the infected wheat tissue usually transitions from green (healthy tissue) to yellow–white (diseased tissue) as the disease progresses. Color has been proven to be the most effective means to distinguish different image objects and realize object recognition among phenotypic traits (e.g., color, texture, and size) extracted from images (Zhang et al., 2018). However, the application of UAV color features in FHB detection has not been explored. Therefore, this study combines color features to further explore the wheat FHB detection methods based on UAV hyperspectral images.

Our study investigated the potential of fusing spectral and image features of UAV hyperspectral images to improve the ability of detecting wheat FHB in the field. The overall technical flow chart is shown in Figure 1. First, we determined the most suitable sensitive spectral features to identify FHB; these features reflect the disease stress of the host. Second, we extracted texture features that could represent the disease distribution based on band images containing the most disease information. Finally, we calculated the color features that characterize disease incidence. We combined multiple algorithms to construct classification models and examine the effect of multi-features on the detection accuracy of FHB. Our goals were to (1) evaluate the performance of UAV hyperspectral images in identifying wheat FHB occurrence; (2) evaluate the potential of multi-features in FHB detection; (3) explore the best classification method for UAV images; and (4) map the occurrence of FHB in a wheat field using the optimal model. In general, we developed a novel method for FHB detection based on UAV images, which forms a basis for the precise prevention and control of FHB.

In the present work, the detailed information contained in UAV hyperspectral images were fully exploited to help identify wheat FHB in the field. FHB can change the pigment, water content, and cell structure of wheat, as well as the structure, shape, and color of the wheat canopy. Therefore, we fused the spectral features that represent internal physiological changes with the image features that represent the spatial information of wheat to effectively detect wheat FHB.

Before analyzing the UAV images, we first evaluated the quality of the UAV hyperspectral images, which is a critical step to ensure that the UAV images accurate identify FHB. We evaluated the quality of the UAV images by comparing and analyzing the data obtained from an ASD spectrometer. The spectral curves of the wheat sample points extracted from the UAV images share a common trend with those of the ASD spectral: a peak and a trough in the VIS–NIR region. However, the values of the ASD spectra were lower than the spectral values obtained from the UAV images, which is likely due to the influence of the bidirectional reflectance distribution function (BRDF) caused by the difference in the geometrical positions of the sun-target-sensors of the two data sets. Some studies have proven that BRDF has a significant impact on UAV hyperspectral data (Burkart et al., 2015). Above 850 nm, the UAV spectral curve shows a downward trend compared with the ASD spectral curve, while the ASD spectral curve has little variation, which is consistent with other scholars’ observations (Gao et al., 2016; Chen et al., 2018). The sensor may have too much noise at the detection boundary. Furthermore, it need cloud-free conditions on the measurement day, so there is a long time interval between the UAV flight and ASD information collection, as well as changes in light conditions. According to Figure 4, the spectral reflectance of sample points obtained by different sensors is significantly correlated within 450–850 nm (R² of 0.99). ASD hyperspectral data are an extensive remote sensing identification method for crop pests and diseases (Cao et al., 2013; Shi et al., 2018; Guo et al., 2020). This spectrum has been used to accurately identify field wheat FHB (Huang et al., 2019a,b; Ma et al., 2020). The high correlation between UAV and ASD spectra further proves the reliability of UAV images.

Next, we extracted the band features, texture features, and color features contained in the hyperspectral images. Table 3 shows the details of the extracted features. The band features we extracted are mainly located in the green edge, red edge, and near-infrared region. The green edge is mainly related to the content of wheat pigments (including carotenoids and chlorophyll) (Al Masri et al., 2017; Zhang Z. P. et al., 2020b), and the position of the red edge is sensitive to the movement of the red edge caused by the change of chlorophyll concentrations (Zhang Z. P. et al., 2020a). Near-infrared wavelengths are primarily related to wheat moisture content, as FHB-infected wheat is accompanied by a temporary increase in transpiration and tissue desiccation (Bauriegel et al., 2011a). The VIS–NIR bands in hyperspectral images are proposed to overcome visual symptom disassociations with DON contamination. Because the DON concentration of wheat is at a low level in the early stages and the typical symptoms of Fusarium damage cannot be detected visually, the spectral features are more conducive to observing the early symptoms of wheat infection (Femenias et al., 2020; Zhang et al., 2021). The GLCM-based texture feature extraction method was based on Fu et al. (2022). The GLCM method describes the texture by studying the spatial correlation characteristics of the gray levels (Haralick and Shanmugam, 1973). In fact, texture information can help distinguish the spatial information independent of tone to identify objects or regions of interest in an image, but it is not recommended to use it by itself due to the poor performance of texture parameters (Sarker and Nichol, 2011). Previously, auxiliary texture information was effectively combined with spectral information to significantly improve the accuracy of wheat GPC estimation (Fu et al., 2022). Therefore, in this study, we attempted to fuse texture and spectral features to improve the detection accuracy of field FHB. The results demonstrate that texture features can serve as complementary information to increase the dimensionality of UAV hyperspectral image data (Table 5). In addition, we calculated some color features by band combinations to indicate different aspects of wheat infection. While texture features may add additional information to FHB estimation, crop infection is more directly related to color information rather than the spatial arrangement of colors (Li et al., 2019). What’s more, since color images highlight specific vegetation greenness and are considered to be less sensitive to changes in light conditions, color features extracted from RGB images have the potential to provide crop growth and nutritional status, immediately providing researchers and farmers with a realistic and intuitive visualization of crop growth status (Du and Noguchi, 2017; Ge et al., 2021). At present, some scholars use color features to estimate the nitrogen density of winter wheat leaves (Rorie et al., 2011), estimate the leaf area index of rice (Li et al., 2019), monitor the growth status of wheat (Du and Noguchi, 2017), and accurately detect wheat FHB at the spikes scale (Huang et al., 2020). However, the effect of FHB detection of field wheat based on color features has not been explored yet. Therefore, in this study, we further supplemented color features in the input models based on spectral and texture features to identify wheat FHB. Actually, color features in UAV digital images are usually based on RGB cameras because UAV systems with RGB cameras are inexpensive, compact, and convenient. In the future, a UAV system suitable for FHB monitoring in the field should be considered. The RGB bands in this study are a basis for future RGB cameras, avoiding the complexity of hyperspectral data processing.

Table 5 shows the model classification results of field wheat with different degrees of infection according to different input variables. The addition of texture and color features can further improve the accuracy of the model compared to methods that use spectral features to detect wheat FHB. As seen in Table 6, the improvement of accuracy is mainly manifested in the difference in the model’s detection of mild, moderate, and severe disease samples. In the early stage of wheat FHB infection (May 3), 57.16% and 42.76% of the field wheat with mild and moderate infection, respectively, could be identified by the model using spectral features. With the addition of texture features and color features, the proportion of mildly infected wheat in the field identified by the model gradually decreased, and the proportion of moderately infected wheat increased. In the late stage of wheat FHB infection (May 8), the model indicated that the proportions of mildly and moderately infected wheat gradually decreased with the addition of image features, and the severely infected area gradually increased. That is to say, before the image features are added, the model always misses the wheat with more severe disease. In fact, as the wheat infection spread, the dry and white areas of the wheat ears became larger until the wheat died (Huang et al., 2020). The addition of image features can enable the model to capture this process. When the information contained in the model increases, the detection of samples with severe disease improves, which is consistent with the results in our study.

The research shows that the fusion of spectral and image features can distinguish the disease incidence of wheat in the field; this method can help future precision agriculture and large area wheat FHB monitoring. However, current research still exists limitations. In addition to considering spectral features, texture features, and color features, some vegetation features, such as Structure Insensitive Pigment Index (SIPI), Anthocyanin Reflectance Index (ARI), Normalized Difference Vegetation Index (NDVI), and Plant Senescence Reflection Index (PSRI), are often used to reflect plant disease stress status (Xiao et al., 2021). Hence, the effectiveness of the vegetation features in wheat FHB detection based on UAV hyperspectral images is worth considering. Additionally, only three machine learning algorithms (RF, BPNN, and SVM) were used in this study. The generalization ability of the models in the temporal and spatial dimensions must be verified. Further consideration can be given to combining data augmentation and deep learning methods to develop more stable and independent models, as well as reduce the uncertainty of model applicability in other regions. Scale has become a popular topic in remote sensing research, and the information contained in a single pixel under different resolutions will change significantly. Appropriate spatial resolution images for agricultural monitoring are needed (Na et al., 2016). Our study only used two images with a spatial resolution of 4 cm to detect FHB, which is relatively simple. Various spatial resolution images are worth considering in the future. Finally, the occurrence of wheat FHB is related to the time of infection (Alisaac et al., 2020) and meteorological factors, such as temperature and humidity. In the future, we will aim to consider wheat infection time and meteorological factors to explore early FHB detection methods and effectively prevent and control FHB occurrence and outbreak. The influence of wheat varieties and the development of various pests and diseases on the spread of wheat FHB cannot be ignored. More researches are needed to investigate the influence of varieties and multiple infections on the model performance in the future.

In this study, the quantitative detection of wheat with mild, moderate, and severe FHB infection in the field was achieved by fusing spectral and image features extracted from the UAV hyperspectral images. After obtaining the hyperspectral images, we first evaluated the quality of the images and identified the data in the 450–850 nm band for subsequent analysis by comparing waveform similarity and correlation with ASD hyperspectral data. Then, we extracted the spectral features that reflect the physiological and biochemical changes within the host, as well as the texture and color features that characterize the spatial changes of wheat. The RF algorithm was used to further eliminate redundant features and improve the operating efficiency of the model. Finally, FHB quantitative detection models, based on different combinations of spectral features, texture features, and color features were formulated by combining BPNN, SVM, and RF algorithms. We evaluated the classification results of the different models, and the FHB-related wheat damage was mapped using the best algorithm. The results show that the spectral features can potentially determine the damage level of FHB, but the performance of the models is not satisfactory. The fusion of spectral features and texture features can improve the model detection level, but the maximum prediction accuracy of the models was only 80%. The model based on the fusion of spectral, texture, and color features was best, and the prediction accuracy of the RF algorithm reached 85%. The damage map illustrates that wheat FHB developed very rapidly over a short time, causing destruction of the crop. This study builds upon previous models in terms of feature types, monitoring methods, and monitoring areas and provides a new methodology for FHB detection in the field by deeply mining features in UAV images and combining multiple spectral advantages.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

HZ: data curation, methodology, and writing-original draft. HM and LL: investigation and data acquisition. JZ and SW: formal analysis, methodology, and supervision. LH: conceptualization and funding acquisition. YD and WH: writing-review and editing. All authors contributed to the article and approved the submitted version.

The work presented here was supported by National Key R&D Program of China (2021YFB3901303), National Natural Science Foundation of China (42071423), Program of Bureau of International Cooperation, Chinese Academy of Sciences (183611KYSB20200080), Alliance of International Science Organizations (Grant No. ANSO-CR-KP-2021-06), Natural Science Research Project of Anhui Provincial Education Department (KJ2019A0030).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.