The glasshouse-based winter wheat experiments were conducted in Jiangsu Province, China, for two consecutive seasons (2019–20 and 2020–21). The hyperspectral reflectance (HR) measurements were performed at the Pailou experiment base of Nanjing Agricultural University (Qinhuai District, Nanjing – 32°1’ N, 118°15’ E), and the fluorescence experiments were conducted at the Intelligent Glasshouse of Nanjing Agricultural University (Xuanwu District, Nanjing – 32°1’ N, 118°12’ E). HR plant material using two wheat varieties (Aikang-58 as susceptible and Sumai-3 as resistant to FHB) was grown successfully in 24 pots (size: 30 cm × 25 cm) in both growing seasons (2019–20 and 2020–21). The detail of the experiment material is given in Table S1 . In each pot, 10 seeds were uniformly grown and managed following the method of Abid et al. (2017), where 12 pots were devoted to each variety and further halved to six for healthy and six for diseased plants. A similar protocol was followed for CFI that was identical to the HR plant material. Whereas, for high throughput phenotyping (HTP), seven wheat cultivars were grown: (1) Bainong-418, (2) Zhongyou-9507, (3) Jimai-31, (4) Wenmai-6, (5) Chianmai-42, (6) Huangpei-R4 as susceptible, and (7) Sumai-3 as resistant to FHB. In total, 56 pots were grown, seven of which were allocated to each cultivar, and five out of seven were inoculated ( Figure 1A ).

The study selected disease sensitive features using Boruta, then after variance inflation factor (VIF) analysis, the partial fusion of selected disease sensitive features (SDSF) was made through ML-SFFS.

Figure 4 shows that both SCC and Pn were severely affected by pathogen infection, but unexpectedly, the study also observed that healthy spikes also showed a slightly decreasing trend. All trends exhibited a noticeable fall in biochemical parameters due to pathogen infestation, but statistically significant differences were not common ( Figure 4A ). A statistically significant difference for Pn appeared at 5% disease percentage (DP) for the period 2019–20, while for the following period (2020–21) it appeared at 6% DP ( Figure 4B ). A similar trend can also be seen in Figure 4C , where a statistically significant difference appeared at 3% DP, while in Figure 4D , it appeared at 4% DP. In nutshell, the pathogen severely affected the biochemical parameters, but SCC were more sensitive than Pn.

Figures 5 , 6 demonstrate the photosynthetic fingerprints of FHB disease invasion on wheat spikes for CFI and HTP, respectively. The DP in respect of days after inoculation (DAI) for two years is shown in the Figure 5A . In Figure 5 , QY showed the clear difference between healthy and diseased samples from 3DAI. Likewise, the Fm_Lss demonstrated the significant difference between healthy and disease spikes, but the F_o showed a balanced response until 5 DAI. However, NPQ responded in absolutely different manner in comparison to all other parameters, it showed first resistance and remained consistent until 5DAI but from 6 to 10 DAI a clear rise in diseased plants was depicted. The HTP shows the clear change (pictorial form-data not shown) in the ears for fv/fm, and CHL.Index ( Figure 6 ).

Figure 7 reveals that the regions of 420–480, 540–680, and 740–860 nm are the spectral regions most highly sensitive to FHB. Moreover, the red-edge (690–730 nm) shift toward the blue region is also prominent, and both the areas and amplitude of the red-edge decreased substantially with the intensification of FHB infestation. Across all mean spectra, there was a gradual increase in the VIS region (400–700 nm), but in the NIR region, there was a continuous decrease. For the first two levels of DS, the NIR region showed an increase, but for the next DS, it decreased substantially.

Figures 8A, B showed that QY_Lss and CHL.Index are more sensitive, respectively. The top five features (CHL.Index, F_v/F_m, QY_Lss, F_m, and QY) were selected with the highest VIP for subsequent analysis as DSF and marked them with red asterisks in Figure 8C . A pooled dataset of both periods (2019–20 and 2020–21) was analyzed and shown in Figure 9C . Although all of the spectral features (SF) in Figures 9A–C showed sensitivity to FHB, the top ten SF (partitioned by the dotted blue line) from each dataset ( Figures 9A–C ) were selected as stable and consistent DSF. Finally, only seven SF (CAR, SRPI, RR, PSSRb, NDVI, PSNDa, and RARSb) were consistent throughout Figures 9A–C (marked by red asterisks), which had shown consistently stable responses to FHB. The resulting DSF showing a VIF of ≤10 were retained as selected disease-specific features (SDSF) ( Table 2 ).

Regarding the feature combination (FC), for the first five levels of DS, the combination was of two to four features but for later ones, only one to two features were sufficient to get the highest overall classification accuracy (CA) ( Table 3 ). These numbers were far below than the multivariate pool of DSF. Although the FC in all three approaches were not identical, some features participated and performed significantly and consistently, i.e., F5 (F_v/F_m) and F8 (NDVI). Figure 10 shows a comparison of the selected features from the SFFS and the use of all SDSF. Although the CA is satisfactory for both approaches, considerable differences prevailed. SVM-SFFS showed better CA than SVM-all, which might be due to a dimensionality factor.

Among all the SDSF, SRPI ( Figure 11C ) yielded the highest RMSE = 17.1 with R ² = 0.86 using an estimated equation developed on the dataset with R ² = 0.88. However, NDVI ( Figure 11D ) showed a lower RMSE = 13.8 with R ² = 0.81, compared to SRPI, R ² between the developed model and the cross-validated datasets exhibited a much greater difference. The minimum RMSE = 9.73 with R ² = 0.86 ( Figure 11E ) during disease estimation was shown by F_v/F_m. Regarding the multivariate models, in respect of their effectiveness in the FHB estimation models, the RF model ( Figure 12A ) resulted in RMSE = 11.11 with R ² = 0.91, the SVM model gave an RMSE = 12.90 with R ² = 0.87, and K-NN outperformed all the others, resulting in an RMSE = 10.20 with R ² = 0.92. Convincingly, all the SDSF explained the significant variation with DS, and model equations had the excellent predictive ability for FHB estimation.

FHB pathogen invasion on wheat spikes damaged the spikelets’ anatomy along with disease proliferation. This damage reduced Pn (Mustafa et al., 2022), SCC, and eventually resulted in the gradual and complete destruction of spike structure. The results confirmed this trend for Pn ( Figures 4A, B ) and SCC ( Figures 4C, D ). Given these, the reduction in the biochemical functions of wheat spikes can be attributed to pathogen development. In addition, SF, CFI and HTP features ( Figures 7 , 8 ) are evident of the photosynthetic damage to the spike structure because these features are pertinent to chlorophyll-related studies ( Table 1 ).

The SDSF from VIP analysis has excellent sensitivity to FHB disease, and each feature clutches specific relevance to plants’ chemistry. For instance, among the SF (SRPI, NDVI, CAR, CHL.Index), SRPI has been previously cited as being most sensitive to chlorophyll and carotenoid components (Gamon et al., 2016). Likewise, NDVI, CAR, and CHL.Index leverage support from the literature as plant pigment indices (Rouse et al., 1974). Previous studies have successfully employed the VI for disease detection in different crops (Ren et al., 2021). Accordingly, FHB detection and monitoring have investigated the VI for hyperspectral imaging and found PSSRa and PSSRb most sensitive (Bauriegel et al., 2011; Mahlein et al., 2019). This study also found these two VI sensitive to FHB disease ( Figure 9 ) but failed to compete with the chosen ones on behalf of the criteria of consistent behavior and VIP. The reason might be that these studies had not adopted consistent features selection approach and claimed correlation-based sensitivity. However, this study selected SRPI, NDVI, CAR, and CHL.Index for classification using a consistent feature selection approach. Resultantly the SF have sensitivity for FHB detection and could be attributed for pigment damage in the plants. The reflectance pattern ( Figure 7 ) and the development of FHB invasion severity are in accord with the findings of previous studies (Ha et al., 2016 and Huang et al., 2020) that also represent the pigment damage and red-edge shift under disease stress.

Chlorophyll fluorescence is a well-known noninvasive approach to examine the photosynthetic fingerprints of stress (biotic or abiotic) on the metabolism of plants (Gitelson et al., 1999; Ehlert and Hincha, 2008; Huang et al., 2019b). Numerous studies have reported Fv/Fm as an integral fluorescence attribute for successful plant examination under applied crop treatments (Harbinson, 2013; Lei et al., 2017). A couple of studies found F_v/F_m from CFI as a strong candidate for FHB detection (Bauriegel et al., 2011; Bauriegel and Herppich, 2014), even when the symptoms were not visible on the glumes (Mahlein et al., 2019). Likewise, in this study, F_v/F_m played a substantial role under VIP analysis for CFI and HTP. Since all the SDSFs were selected from the plant pigment-related studies, these could be potential candidates for studying FHB fingerprints on wheat spikes for pigment damage and detection.

In contrast to ML-SFFS, a few studies have examined optimal features or feature fusion on disease detection (Zarco-Tejada et al., 2018). Ultimately, feature selection employing either approach is an effective tactic for handling the large volume of data by reducing redundant information. Hence, ML-SFFS following VIF analysis easily overcomes the collinearity challenge and also deals with the computation load (Tian et al., 2021). In former investigation, Mahlein et al. (2019) obtained no significant improvement with the sensor fusion approach after three days of disease inoculation. However, Bauriegel et al. (2011) claimed improved CA for hyperspectral and fluorescence imaging fusion. In current study at the asymptomatic scale obtained high CA and at DS1 it manifested 87% CA. Three to four features could claim the highest potential CA, which is interpreted as each feature showing variation under pathogen attack, and the overall obtained accuracy was satisfactory. The notable factor is that fluorescence features competed strongly with VI at each level of DS. In fact, over the range DS1–8, F5 shared in each classification approach (RF, K-NN, and SVM), except at DS3 in K-NN. Moreover, this ML methodology explains the interpretability and rationality of the FC, because some features might perform better at one DS than at another. For example, F5 intervened in most levels of DS compared to any other feature due to its great sensitivity to FHB (Mahlein et al., 2019) in the studied datasets. The inclusion of different features at different levels of DS also help to interpret the disease-specific responses to the specific features. For example, all spectral features showed sensitivity to FHB in the VIP algorithmic test, but few (DSF) were of more importance where further redundancy led to obtain the SDSF. However, by performing SDSF and ML-SFFS maneuvers, the most relevant features for studying the photosynthetic fingerprints of FHB for classification were selected and abundant redundancy was filtered out. Similar approaches have been adopted to determine the effective plant traits in Xylella fastidiosa infection (Zarco-Tejada et al., 2018).

The SDSF adjusted the combination of different features for FHB classification at different levels of DS and resulted in the best CA under ML-SFFS. Subsequently, employed the SDSF for FHB estimation by feeding into univariate and multivariate estimation modeling. Both results are examinable for proxy estimation of FHB. In comparison, multivariate estimation resulted in better accuracy than univariate models ( Figure 11 , Figure 12 ), which agrees with Zhang et al. (2014), who estimated yellow rust in wheat using wavelet features and VI.

Extraction and selection of features from hyperspectral and chlorophyll fluorescence data can significantly enhance computing efficiency and highlight the essential elements for the development of classification methodologies. In contrast, feature selection algorithms have been demonstrated to be efficient for maintaining important information while reducing computation time (Huang et al., 2019a). The suggested ML-SFFS classification approach outperforms earlier disease classification models by combination of sensitive features for high CA. The significant rise in CA with DS shows that FC with two to four features could give a higher CA than all DSF with lower computational cost ( Figure 10 ). Individual spectral features (Mahlein et al., 2013; Shi et al., 2018) and ML (Rumpf et al., 2010) have been used in previous studies with promising results for the detection of plant diseases (Cheng et al., 2010; Mahlein et al., 2013). However, the majority of previous studies utilized complex classification algorithms for disease detection and only a few attempts were made to enhance FC for higher classification performance (Rumpf et al., 2010; Zarco-Tejada et al., 2018). By combination of sensitive features, the weak features to disease stress at the earliest stage of infection could be successfully amplified. In this work, the application of VIF analysis and ML-SFFS algorithm enabled not only the decrease of collinearity among predictor variables, but also the reduction of computational burden.

ML-based classification ensures logic and interpretability of FC picked from SDSF by the SFFS technique. For instance, reflectance and fluorescence data characteristics may have performed well but could be uninterpretable and case-specific. However, all FC identified from SDSF were directly connected to FHB infection, physiological and morphological changes in infected spikes, allowing this methodology’s generalization and transferability to examine other diseases (Zarco-Tejada et al., 2018). Sensitive features for disease detection have been discussed in the literature recently (Poblete et al., 2020). A FC usually incorporate numerous plant attributes, which better illustrates FHB’ infection’s complicated physiological processes, and can explain this variance through spectral and fluorescence features.

The photosynthetic fingerprints, particularly on SCC were more sensitive than Pn under FHB pathogen invasion while several studies have demonstrated its integral role in grain filling (Tambussi et al., 2007; Jia et al., 2015). This might facilitate FHB detection on a large scale that could be challenging in the context of destructive sampling. In addition, studies have explored the VI for FHB detection (Bauriegel et al., 2011; Bauriegel and Herppich, 2014), which can be practiced with the currently available technologies. Moreover, numerous studies have resulted in efficient disease detection for other crops deploying VI (Zarco-Tejada et al., 2018; Mahlein et al., 2019; Tian et al., 2021). Despite attaining high CA at different scales, the ML-SFFS approach has revealed some key concerns over FC, the combination of different features at each scale and the inclusion of different sensor data. Consequently, this can restrict its large-scale application for disease detection because relative disease sensitivity can vary at different levels of DS. However, for disease quantification and estimation, the SDSF exhibited substantial potential for univariate and multivariate modeling. Moreover, SDSF can be employed in remotely sensed disease detection systems at different scales for deep phenotyping of wheat spikes. Hence, this disease detection methodology can be applied in different farm fields developing a manageable data acquisition setup.