Development of Spectral Disease Indices for Southern Corn Rust Detection and Severity Classiﬁcation

: Southern Corn Rust (SCR) is one of the most destructive diseases in corn production, signiﬁcantly a ﬀ ecting corn quality and yields globally. Field-based fast, nondestructive diagnosis of SCR is critical for smart agriculture applications to reduce pesticide use and ensure food safety. The development of spectral disease indices (SDIs), leaf spectral signatures assist in the accurate diagnosis of SCR, no SDIs-based model has been reported for the ﬁeld-based SCR monitoring. Here, to we developed SDIs-based monitoring models to detect SCR-infected leaves and classify SCR damage severity. In detail, we ﬁrst collected in situ leaf reﬂectance spectra (350–2500 nm) of healthy and infected corn plants with three severity levels (light, medium, and severe) using a portable spectrometer. Then, the RELIEF-F algorithm was performed to select the most discriminative features (wavelengths) and two band normalized di ﬀ erences for developing SDIs (i.e., health index and severity index) in SCR detection and severity classiﬁcation, respectively. The leaf reﬂectance spectra, most sensitive to SCR detection and severity classiﬁcation, were found in the 572 nm, 766 nm, and 1445 nm wavelength and 575 nm, 640 nm, and 1670 nm wavelength, respectively. These spectral features were associated with leaf pigment and leaf water content. Finally, by employing a support vector machine (SVM), the performances of developed SCR-SDIs were assessed and compared with 38 stress-related vegetation indices (VIs) identiﬁed in the literature. The SDIs-based models developed in this study achieved an overall accuracy of 87% and 70% in SCR detection and severity classiﬁcation, 1.1% and 8.3% higher than the other best VIs-based model under study, respectively. tool fast, ﬁeld our ﬁrst Z.L., G.C., and B.X.; resources, R.M. and J.Y.; writing—original draft preparation R.M. and Z.L.; writing—review and editing, R.M., Z.L., J.Y., G.C., F.Z., L.Z., and B.X.; funding acquisition, R.M.


Introduction
Southern Corn Rust (SCR), caused by Puccinia polysora Underw, is a foliar disease, significantly affecting corn quality and yield at global scales [1,2]. In epidemic years, SCR can reduce corn production by up to 50% or even 100% [2]. Especially during recent years, the annual corn yield loss caused by SCR has increased sharply because of elevated winter minimum temperatures and the lack of SCR-resistant corn cultivars [2,3]. For example, it is estimated that in China, one of the countries most affected by SCR, the corn yield loss caused by SCR in 2015 was 756 million kg, which was as high as 8.8-times that of the annual average of 2008-2014 [2,3]. Also, in 2015, the corn yield loss caused by SCR was estimated to be 3.193 billion kg in the United States and Canada, ranking as the sixth most serious disease in terms of yield loss [4].
To control disease spread and reduce yield loss, spatial-explicit information on the disease-infected plants is indispensable in precision crop protection for guiding pesticide use and other management activities [5][6][7]. The development of a fast, nondestructive detection method of plant diseases (e.g., SCR) over large areas is thus a demanding challenge [6,7]. Visual inspection is the traditional method for detecting plant diseases in the field, but it is inefficient and error-prone. During recent years, remote sensing technology has shown great potential in the rapid and accurate detection of plant diseases with various crops in a nondestructive way [6,7].
Reflectance spectroscopy has been proven to be an effective approach to distinguish disease-infected leaves from healthy leaves in a fast and nondestructive way [6,[8][9][10]. This approach utilizes changes in leaf optical properties impacted by plant disease at narrow-band wavelengths to detect plant diseases [7]. Changes in reflectance at visible wavelengths (400-700 nm) are mainly determined by pigment concentration and photosynthesis efficiency, while near-infrared (NIR, 700-1100 nm) wavelengths are strongly influenced by leaf internal structure, and shortwave infrared spectra (SWIR, 1100-2500 nm) are usually impacted by leaf water content and other biochemical compositions. Reflectance spectroscopy has been successfully applied in numerous studies for plant disease detections, such as potato late blight [8,11], rice glume blight disease [12], wheat yellow rust disease [13][14][15], maize dwarf mosaic and Helminthosporium maydis disease [16], tomato leaf diseases (e.g., late blight, target and bacterial spots) [17,18], and so on. However, the processing and analysis of reflectance spectra data with large number of narrow-band wavelengths is still complex and time-consuming [19,20].
Previous studies have indicated that some specific spectral features (wavelengths) can provide significantly higher discrimination power than others [6,7,9]. Thus, to further improve and simplify plant disease detection, researchers have developed spectral disease indices (SDIs) based on certain spectral features highly correlated to changes caused by a specific disease [9]. Using the RELIEF-F algorithm, Mahlein et al. [9] first identified the most important spectral features, and then developed SDIs most sensitive for detecting three types of leaf diseases of sugar beet (i.e., Cercospora leaf spot, sugar beet rust, and powdery mildew) based on leaf reflectance spectra. Furthermore, Bauriegel et al. [20] designed a head blight index (HBI) for detecting wheat ears infected by fusarium head blight at the early stage in the field, and results indicated that the HBI-based classification method was a promising alternative of Spectral Angle Mapper-based method using the full spectral wavelength under study, as the latter was more complicated and time-consuming. In addition, Huang et al. [21] found that photochemical reflectance index (PRI) can effectively quantify the severity of yellow rust in wheat using both in situ canopy reflectance and airborne hyperspectral measurements with a coefficient of determination (R2) of > 0.91. Similarly, Zheng et al., [22] developed optimal spectral indices for detecting yellow rust in wheat at different growth stages, based on canopy spectral reflectance measurements and the classification accuracies of all VIs-based models were above 80% in their research. Therefore, the SDIs could potentially be a promising tool for the fast, nondestructive diagnosis of SCR in the field. However, little is known about the spectral signatures of SCR-infected leaves with different severity levels, and no SDIs-based model has been reported for detecting SCR or classifying SCR severity using in situ leaf reflectance spectra. Hence, in this study, to fill in those knowledge gaps, we proposed and evaluated SDIs-based models for detecting SCR-infected leaves and classifying their damage severity. Based on previous studies [9,23], SCR-SDIs were developed by combining single and a normalized wavelength difference from the pool of most important spectral features for the SCR monitoring. Specifically, the main objectives of this paper were to (I) characterize leaf-level spectral signatures of corn leaves infected by SCR; (II) determine the spectral features and normalized differences most important for developing SDIs in SCR detection and SCR damage severity classification, respectively; and (III) evaluate and compare the performances of our developed SCR-SDIs with the 38 stress-related vegetation indices (VIs) in the literature.

Study Site
The study site is in Huazhong Agricultural University Experimental Station (18.38N, 109.18E) in Sanya, Hainan province, China ( Figure 1). This site is locaed in the tropical monsoon climate zone, with an average annual temperature and precipitation rate of 25.5 • C and 1500 mm, respectively. Since first reported in 1972, SCR became an epidemic in Hainan annually [24]. Hainan is thus an ideal place for SCR-related studies. Based on our multiple years of planting experience, the flowering period is the critical phenological cycle for SCR monitoring in our study site. About two weeks after pollination in early January 2020, an experiment field with a natural and widespread outbreak of SCR was identified in the study site. Based on visual inspection, we did not find other major stress or disease other than the target disease, and no pesticide was applied in the experiment field. The experiment field included dozens of corn cultivars, each of which had 10 plants in total, respectively. Standard agricultural practices were conducted for the crop under study with the same water and fertilizer application throughout the growing season.
Remote Sens. 2020, 12, 3 of 16 Therefore, the SDIs could potentially be a promising tool for the fast, nondestructive diagnosis of SCR in the field. However, little is known about the spectral signatures of SCR-infected leaves with different severity levels, and no SDIs-based model has been reported for detecting SCR or classifying SCR severity using in situ leaf reflectance spectra. Hence, in this study, to fill in those knowledge gaps, we proposed and evaluated SDIs-based models for detecting SCR-infected leaves and classifying their damage severity. Based on previous studies [9,23], SCR-SDIs were developed by combining single and a normalized wavelength difference from the pool of most important spectral features for the SCR monitoring. Specifically, the main objectives of this paper were to (I) characterize leaf-level spectral signatures of corn leaves infected by SCR; (II) determine the spectral features and normalized differences most important for developing SDIs in SCR detection and SCR damage severity classification, respectively; and (III) evaluate and compare the performances of our developed SCR-SDIs with the 38 stress-related vegetation indices (VIs) in the literature.

Study Site
The study site is in Huazhong Agricultural University Experimental Station (18.38N, 109.18E) in Sanya, Hainan province, China ( Figure 1). This site is locaed in the tropical monsoon climate zone, with an average annual temperature and precipitation rate of 25.5°C and 1,500 mm, respectively. Since first reported in 1972, SCR became an epidemic in Hainan annually [24]. Hainan is thus an ideal place for SCR-related studies. Based on our multiple years of planting experience, the flowering period is the critical phenological cycle for SCR monitoring in our study site. About two weeks after pollination in early January 2020, an experiment field with a natural and widespread outbreak of SCR was identified in the study site. Based on visual inspection, we did not find other major stress or disease other than the target disease, and no pesticide was applied in the experiment field. The experiment field included dozens of corn cultivars, each of which had 10 plants in total, respectively. Standard agricultural practices were conducted for the crop under study with the same water and fertilizer application throughout the growing season.

Methods
To monitor SCR using in situ leaf reflectance spectra, we took the following three steps ( Figure 2). First, the spectra and photos of healthy and SCR-infected leaves with three levels of severity were collected in the field and pre-processed in the lab. Second, using the RELIEF-F algorithm, two SDIs for SCR detection (i.e., Healthy Index, HI) and severity classification (i.e., Severity Index, SI) were developed, respectively. Finally, based on our developed SDIs and the other 38 stress-related VIs in the literature (Table S1), different models for SCR detection and severity classification were established and compared using the Support Vector Machine (SVM) algorithm.

Methods
To monitor SCR using in situ leaf reflectance spectra, we took the following three steps ( Figure  2). First, the spectra and photos of healthy and SCR-infected leaves with three levels of severity were collected in the field and pre-processed in the lab. Second, using the RELIEF-F algorithm, two SDIs for SCR detection (i.e., Healthy Index, HI) and severity classification (i.e., Severity Index, SI) were developed, respectively. Finally, based on our developed SDIs and the other 38 stress-related VIs in the literature (Table S1), different models for SCR detection and severity classification were established and compared using the Support Vector Machine (SVM) algorithm.

Data Acquisition and Preprocessing
We measured and recorded the healthy and SCR-infected leaves with different severity levels in the field after visual inspection. Specifically, to reduce the effects of leaf age [25,26], only the last leaf from the bottom was selected, and the selected leaves were also visually checked to make sure there was no other damage or stress than SCR. Following the procedure outlined in previous corn foliar disease studies [27,28], SCR severities were visually rated and classified into three levels based on the percentages of lesion areas on leaves by the help of a plant pathologist ( Figure 3 and Table 1). Digital photographs of leaf samples were taken in the field for further reference.

Data Acquisition and Preprocessing
We measured and recorded the healthy and SCR-infected leaves with different severity levels in the field after visual inspection. Specifically, to reduce the effects of leaf age [25,26], only the last leaf from the bottom was selected, and the selected leaves were also visually checked to make sure there was no other damage or stress than SCR. Following the procedure outlined in previous corn foliar disease studies [27,28], SCR severities were visually rated and classified into three levels based on the percentages of lesion areas on leaves by the help of a plant pathologist ( Figure 3 and Table 1). Digital photographs of leaf samples were taken in the field for further reference.

Methods
To monitor SCR using in situ leaf reflectance spectra, we took the following three steps ( Figure  2). First, the spectra and photos of healthy and SCR-infected leaves with three levels of severity were collected in the field and pre-processed in the lab. Second, using the RELIEF-F algorithm, two SDIs for SCR detection (i.e., Healthy Index, HI) and severity classification (i.e., Severity Index, SI) were developed, respectively. Finally, based on our developed SDIs and the other 38 stress-related VIs in the literature (Table S1), different models for SCR detection and severity classification were established and compared using the Support Vector Machine (SVM) algorithm.

Data Acquisition and Preprocessing
We measured and recorded the healthy and SCR-infected leaves with different severity levels in the field after visual inspection. Specifically, to reduce the effects of leaf age [25,26], only the last leaf from the bottom was selected, and the selected leaves were also visually checked to make sure there was no other damage or stress than SCR. Following the procedure outlined in previous corn foliar disease studies [27,28], SCR severities were visually rated and classified into three levels based on the percentages of lesion areas on leaves by the help of a plant pathologist ( Figure 3 and Table 1). Digital photographs of leaf samples were taken in the field for further reference.   Finally, to reduce the potential effects of phenotype difference, 54 leaf samples from eight corn cultivars (Jing724, Jing92, Chang7-2, PH6WC, DH382, DH351, 8H534, 8H953), each of which had 6-8 leaf samples selected, of which at least one leaf sample belonged to each of three infection levels, were used for this study (Table S2).
The in situ spectral reflectance of corn leaf samples was measured using a Spectra Vista Corporation (SVC) HR-1024i spectroradiometer with a leaf clip holder attached to a pistol grip foreoptic (SVC, Poughkeepsie, NY, USA) in the daytime. The SVC spectroradiometer had a spectral range of 350-2500 nm with varied spectral resolutions (3.5 nm at 700 nm, 9.5 nm at 1500 nm, and 6.5 nm at 2100 nm). To remove the noises at the edges of spectra, only spectra between 400 nm and 2450 nm were kept and analyzed in our study. Based on a Gaussian function, the collected spectra were resampled to 1 nm for instrument output. The pistol grip foreoptic had an internal, calibrated halogen light source to illuminate the samples during leaf reflectance measurements under field conditions. For each leaf sample, reflectance spectra were measured on the upper-middle-lower parts of the leaf adaxial surface six times, and then were averaged to represent the optical properties of the whole leaf for final data analysis ( Figure 3). To ensure data quality, the spectroradiometer was preheated for 20 min prior to measuring, and white reference measurements were made every 5 min using a spectralon panel (Labsphere, North Sutton, NH, USA).

Feature Selection and Combination
Feature selection is necessary before the development of SCR-specific SDIs, as irrelevant features (wavelengths) can increase the computation task and lead to wrong results [23,29]. There are multiple algorithms available for feature selection [25]. As the RELIEF-F algorithm is one of most classical ones and has been widely applied in the field of SDIs development, we chose it to extract the wavelengths most important for SCR detection and severity classification [9,23]. The RELIEF-F algorithm can estimate discriminative power of certain features in multiclass classification problems with incomplete and noisy data in terms of their performances in separating samples of different classes close to each other [23,30]. The RELIEF-F algorithm has been widely used in the development of SDIs for detecting plant diseases [9,23]. A weight for a feature of a measurement vector was calculated by the RELIEF-F algorithm to quantify feature relevance [23]. Generally, the RELIEF-F algorithm consists of two steps: (1) Search two nearest neighbors of the same class (hit) and from the different classes (miss) for a given number of samples in the neighborhood; (2) Calculate the weight (relevance) of each wavelength (feature) by the sum of the Euclidean distance between nearest misses and nearest hits for all samples [9,23]. Finally, wavelengths with high weight values at different spectral regions (i.e., weight values larger than 0.025-nm at 50-nm wavelength intervals) were chosen as the candidate spectral features for developing SCR-specific SDIs in our study. Before feature selection with the RELIEF-F algorithm, Savitzky-Golay smoothing was conducted to filter spectral signal noises from all samples. Savitzky-Golay smoothing and the RELIEF-F algorithm were both performed in the R environment (version 3.5.3; https://www.r-project.org/) with the packages "hsdar" and "CORElearn," respectively.
The optimized combinations of individual wavelength and a normalized wavelength difference spectral index (NDSI) (i.e., SDIs) were further selected for detecting SCR-infected leaves and classifying SCR severity, respectively [9]. In other words, we developed two SDIs according to the previous feature selection results by the RELIEF-F algorithm: The HI for discriminating healthy and SCR-infected leaves Remote Sens. 2020, 12, 3233 6 of 16 and the SI for classifying severity of infected leaves. The formula for SDIs development is listed as follows, according to previous studies [9]: where a, b, and c (a b c) are wavelengths chosen from the pool of selected wavelengths, and d is the coefficient with a value range of (−1, +1) at a step of 0.1. Here, the HI and SI were finally determined using an exhaustive search strategy [9].

Performances of Different Spectral Indices in SCR Detection and Severity Classification
In this study, a machine learning algorithm (i.e., SVM) was chosen to build SCR detection and severity classification model, based on the developed HI and SI, respectively. The SVM algorithm, a supervised learning method, has been widely used for detecting and classifying crop diseases [8,20]. Following previous similar studies [15,31], the radial basis function (RBF) was used as the kernel function of our SVM model for class separation, and the cost of constraints (C) and sigma (σ)-two key parameters of an SVM classifier-were determined by a grid-based search strategy [32].
A 10-fold cross validation strategy was used to evaluate and compare the performances of the developed models. Specifically, the samples were randomly split into two parts for training the SVM classifier (90%) and validating the classification accuracy (10%), respectively. This process was repeated 10 times. At last, in addition to the Overall Accuracy (OA) and Macro F1-scores of both models, the sensitivity and specificity of SCR detection model and the balanced accuracy of SCR severity model were calculated and reported (Equations (2)- (7)).
Balanced accuracy = (Sensitivity + Speci f icity)/2 (6) where TP (true positive), FP (false positive), TN (true negative), and FN (false negative) are four values computed based on classification confusion matrix, and N is the number of class types under study. In addition, the capability of SDIs in SCR detection and severity classification was compared with a total of 38 stress-related VIs in the literature (Table S1). These VIs can be grouped into three types: Simple ration indices, red edge parameters, and normalized ratio indices. Due to the large number of VIs under study, we just listed the first five most discriminative VIs in SCR detection and severity classification, respectively (Table 2). To see the whole set of implemented VIs, please refer to Table S1 in the Supplementary Materials.  SRI Simple ratio index R 515 /R 550 Hernández-Clemente et al., [41] This table only includes the first five most optimal spectral indices for SCR detection and severity classification, selected from 38 stress-related spectral indices in the literature. In formula, D: The first derivative of spectra; R: The base wavelengths; ∆: The integration width.

Spectral Signatures of Healthy and SCR-Infected Corn Leaves
The mean spectra of healthy leaf and each severity group of SCR-infected leaves with standard deviation are calculated and plotted in Figure 4. In general, the leaf reflectance spectra were sensitive to SCR infection, with relatively large spectra variations at the green, red-edge, NIR, and SWIR wavelengths. Specifically, NIR reflectance, associated with leaf internal structure, decreased with severity. SWIR reflectance, associated with leaf water content and other leaf biochemical compositions, also decreased with severity. On the contrary, the reflectance at the green (550 nm) to red (680 nm) and red-edge regions (670-760 nm), most controlled by leaf chlorophyll content and photosynthesis efficiency, increased with severity.

Spectral Signatures of Healthy and SCR-Infected Corn Leaves
The mean spectra of healthy leaf and each severity group of SCR-infected leaves with standard deviation are calculated and plotted in Figure 4. In general, the leaf reflectance spectra were sensitive to SCR infection, with relatively large spectra variations at the green, red-edge, NIR, and SWIR wavelengths. Specifically, NIR reflectance, associated with leaf internal structure, decreased with severity. SWIR reflectance, associated with leaf water content and other leaf biochemical compositions, also decreased with severity. On the contrary, the reflectance at the green (550 nm) to red (680 nm) and red-edge regions (670-760 nm), most controlled by leaf chlorophyll content and photosynthesis efficiency, increased with severity.

SDIs for SCR Detection and Severity Classification
Based on the weight values calculated by RELIEF-F algorithm, eight discriminative wavelengths were selected to develop SDIs for detecting SCR-infected leaves (i.e., HI) and classifying SCR severity levels (i.e., SI) ( Table 3 and Figure 5). There were spectral wavelengths with relatively high weight values (i.e., high discriminative power) at the green, red, red-edge, NIR and SWIR regions, consistent with our qualitative analysis on the mean spectra of healthy and SCR-infected leaves with different severity levels ( Figure 4). In general, the weight values of SCR detection were mostly larger than that of SCR severity classification, but the shapes of their weight value plots were similar ( Figure 5). Consequently, the 1445 nm, 572 nm, and 766 nm wavelengths with a coefficient value of 0.2 were selected to develop HI, and the 575 nm, 1670 nm, and 640 nm wavelengths with a coefficient value of -1 were used for developing SI (Equations. (8,9)).

SDIs for SCR Detection and Severity Classification
Based on the weight values calculated by RELIEF-F algorithm, eight discriminative wavelengths were selected to develop SDIs for detecting SCR-infected leaves (i.e., HI) and classifying SCR severity levels (i.e., SI) ( Table 3 and Figure 5). There were spectral wavelengths with relatively high weight values (i.e., high discriminative power) at the green, red, red-edge, NIR and SWIR regions, consistent with our qualitative analysis on the mean spectra of healthy and SCR-infected leaves with different severity levels ( Figure 4). In general, the weight values of SCR detection were mostly larger than that of SCR severity classification, but the shapes of their weight value plots were similar ( Figure 5).

Performances of the SCR-Specific Indices
The performance of our developed HI for detecting SCR-infected leaves was shown and compared with 38 stress-related VIs in the literature (Table S1). The validation results of the first five most discriminative VIs, among the 38 published stress-related VIs under study, are listed for detecting SCR-infected leaves in Table 4 and Figure S1 (please refer to Table S1 in the supplementary materials for the whole list). In general, the performance of HI model was the best for detecting SCRinfected leaves. The sensitivity (0.833), OA (87%), and Marco-F1 (0.856) values of the HI model were all highest among VIs under study, and the specificity value (0.889) of the HI model ranked second. The D730/D706-model also had a higher value in sensitivity (0.950), but the model specificity was much lower (0.678) compared to that of the HI model. Similarly, the performance of our developed SI for classifying SCR severity is shown and compared with the other VIs under study in Table 5 and Figure S2. In general, the accuracy of SCR severity model was lower than that of SCR detection model, but the overall performance of our developed SI was still the best among VIs under study for classifying SCR severity levels. The OA (70%), Marco-F1 score (0.698), and balance accuracies of the light (0.78), medium (0.778), and severe (0.767) level were all highest. Consequently, the 1445 nm, 572 nm, and 766 nm wavelengths with a coefficient value of 0.2 were selected to develop HI, and the 575 nm, 1670 nm, and 640 nm wavelengths with a coefficient value of −1 were used for developing SI (Equations (8) and (9)).

Performances of the SCR-Specific Indices
The performance of our developed HI for detecting SCR-infected leaves was shown and compared with 38 stress-related VIs in the literature (Table S1). The validation results of the first five most discriminative VIs, among the 38 published stress-related VIs under study, are listed for detecting SCR-infected leaves in Table 4 and Figure S1 (please refer to Table S1 in the supplementary materials for the whole list). In general, the performance of HI model was the best for detecting SCR-infected leaves. The sensitivity (0.833), OA (87%), and Marco-F1 (0.856) values of the HI model were all highest among VIs under study, and the specificity value (0.889) of the HI model ranked second. The D730/D706-model also had a higher value in sensitivity (0.950), but the model specificity was much lower (0.678) compared to that of the HI model. Similarly, the performance of our developed SI for classifying SCR severity is shown and compared with the other VIs under study in Table 5 and Figure S2. In general, the accuracy of SCR severity model was lower than that of SCR detection model, but the overall performance of our developed SI was still the best among VIs under study for classifying SCR severity levels. The OA (70%), Marco-F1 score (0.698), and balance accuracies of the light (0.78), medium (0.778), and severe (0.767) level were all highest.

Discussion
As stated before, the SDIs-based method for plant disease detection is computationally simple, fast, and nondestructive [9,20,23,42], which is critical for efficiently instructing the pesticide use and other management activities in crop protection in the field. This is the first study, to our knowledge, to adopt a SDIs-based (i.e., HI and SI) method for detecting SCR-infected leaves and classifying SCR damage severity. The validation results indicated that our developed SDIs outperformed 38 stress-related VIs commonly found in literature (Table S1) in terms of both SCR detection and damage severity classification (Tables 4 and 5). Our study thus highlights a fast and nondestructive method for SCR monitoring in the field compared with other techniques, like polymerase chain reaction (PCR)-based diagnostics [43,44].

Spectral Signatures of SCR-Infected Leaves
Foliar pathogens can change leaf biochemical and biophysical contents (e.g., leaf pigment, leaf water content, and leaf internal structure), and these changes can affect the leaf spectral signatures, which are the basis of the remote sensing of plant diseases [7,17,45,46]. In our study, the main influence of SCR to leaf reflectance was found in the visible (VIS) wavelength from 550 m to 700 nm, in the red-edge-NIR wavelength from 700 nm to 850 nm, and in the SWIR wavelength from 1300 nm to 1800 nm, respectively (Figures 4 and 5). The mentioned wavelengths were associated with leaf chlorophyll content, leaf water content, leaf photosynthetic efficiency, and leaf internal structure [12,22]. In addition, there were correlations between SCR damage severity and changes in VIS, red-edge, NIR, and SWIR spectral regions, making SCR severity classification possible. As the SCR severity increased, the leaf reflectance at VIS-red edge regions increased but the leaf reflectance at the NIR-SWIR regions decreased (Figures 4 and 5), which is consistent with previous studies on the remote sensing of crop disease and pests at the leaf level [9,20,23,42]. It also indicated that the SI developed here could be potentially used for quantifying the resistance of corn variety to SCR, like the phenotyping assessment of crop diseases for sugar beet and wheat [47,48].
Unlike most of the previously published similar studies on developing SDIs [9,20,23,42] which have focused on the VIS-NIR spectral range, SWIR regions were included in our study as candidate features for developing SDIs, and the feature selection results confirmed the importance of SWIR regions in both SCR detection and severity classification (Table 3 and Figure 5). Therefore, more attention should be paid to the SWIR-related VIs or features for the remote sensing of crop diseases and pests in the future. However, there is a tradeoff between performance improvement using SWIR-involved new indices and the cost increase using the equipment with SWIR measurement capability in application, as the spectrometer with a range of 400-2500 nm is considerably more expensive than that with a range of 400-1000 nm.
We found that some of the previously published VIs (e.g., D730/D706 and DDI), sensitive to leaf chlorophyll content [33,34], had good performances in detecting SCR-infected leaves (Tables 3  and 4). When classifying severity levels for SCR damage, previously developed VIs (e.g., DWSI and PRI), sensitive to water stress [38,39], showed good results (Tables 3 and 5). This phenomenon could be explained by the fact that the SCR fungus (i.e., Puccinia polysora Underw) can directly generate uredium on the leaf surface, damage leaf chlorophyllous tissues, and cause leaf necrosis until the plant eventually dies [49]. In summary, it is suggested that changes in leaf chlorophyll content and leaf water content are the basis of spectral signatures of SCR.

Advantages of SDIs-Based Method for Plant Disease Monitoring
Compared to other new techniques, the SDIs-based method has several advantages that make it powerful for plant disease monitoring in the field. First, it is simple to compute. The SDIs were constructed in simple forms for easy calculations and the corresponding developed models only required light computations for implementation. In addition, no expert knowledge is required, such as the PCR-based method [7], to implement it. Second, it is fast, nondestructive, and takes only a few seconds to scan. Thus, it is possible to make repeat measurements of the same leaves and plants through their life cycle without negative effects on the crop yield, which is essential for large-scale plant disease monitoring in production. The fast screening of hundreds of plants in a short time also enables plant disease monitoring in a timely manner, which is crucial for taking appropriate measures in precision crop protection [5,50,51] and for the phenotyping of crop disease resistance in crop breeding [47,48]. Third, the SDIs-based method can make field-deployable spectra instruments for plant disease monitoring more commercially available [52,53]. A portable spectroradiometer with full-range (i.e., VIS-NIR-SWIR) measurement capability, such as the one used in this study, is expensive (~tens of thousands of US dollars), but an instrument with only certain spectra wavelengths measurement capability is much cheaper. A good example is the SPAD meter (~thousands of US dollars) for crop nitrogen estimation, which is also based on the leaf spectral measurement of certain wavelengths [54]. Our SCR detection model obtained an OA of 87% using our developed SDIs, which is comparable to previous similar studies using leaf spectral features for crop disease detection [9,13,15]. On the other hand, using the same methodology, our SCR severity model obtained an OA of 70%, which is relatively lower. The SCR detection and severity classification errors come from the fact that the reflectance standard deviations for each of the different spectral bands were relatively high and overlapped between healthy and infected leaves and among different severity classes. It is difficult to locate the bands sensitive to SCR infection and damage levels just based on visual inspection. The RELIEF-F algorithm was thus applied to locate the bands sensitive to SCR detection and severity classification, respectively (Table 3). Then, based on the selected sensitive bands, two models were developed for the rapid identification of the SCR and its severity (Tables 4 and 5). Again, the varying performances of these two kinds of model were consistent with the differences in discriminative power of spectral features (wavelengths) in SCR detection and severity classification ( Figure 5). This phenomenon reflects the fact that the spectral differences between healthy and SCR-infected leaves in this study were much larger than that among SCR-infected leaves with different severity levels (Figures 4 and 5).

Limitations and Future Studies
We acknowledge that some limitations exist in this study. First, we did not evaluate our proposed method for SCR detection by considering other types of crop pest and disease, which may cause similar spectral responses of SCR. However, previous studies indicate that different crop pest and disease have their unique spectral signatures, which could be used for differentiating each other [5,9]. Second, due to the relatively small datasets, we did not evaluate the robustness and effectiveness of our developed SDIs at different maize growing stages or other study sites. However, previous similar studies [10,19,29] in spectroscopic analysis of plant pest and disease also only used dozens of plant samples for their research. In addition, our study still serves as the one of first few efforts to explore unique spectral signatures for detecting SCR-infected leaves and classifying SCR damage severity in a rapid and nondestructive way. Thus, our study is a critical basis for remote sensing of SCR in wall-to-wall precision crop protection applications over large areas [21,55].
As an epidemic disease, SCR occurs in patches at the field scale. The fast development of unmanned aerial vehicles (UAVs) with hyperspectral capability provides an unprecedented opportunity for monitoring SCR at the field scale [56][57][58]. However, monitoring SCR-infected plants at the canopy level is more complicated, as the effects of canopy structure, inhomogeneous illumination conditions, leaf ages, backgrounds, and solar-view geometry to canopy spectra should be considered [25,45]. Thus, the developed SDIs-based method cannot be directly applied to UAVs-acquired hyperspectral measurements for monitoring SCR at the field scale. VI combinations (e.g., VI ratio and sum) have been shown effective to reduce the effect of the canopy structure on remote sensing of vegetation biochemical parameters [59,60]. The imaging spectroscopy technique provides a possibility to focus on disease-infected, sunny foliar areas, regardless of background materials and solar-view geometry [25,61]. More studies should thus be conducted to explore the sensitivity of various VI pairs, including our developed SDIs, to canopy structure for monitoring plant disease and pest over large areas, and a new method using imaging spectroscopy is also needed to remove the background effects in the field considering the geometry view.
Previous studies have indicated that the combined use of spectral and textural features of imaging spectroscopy can significantly improve the accuracy in crop disease detection [15]. As we used a point spectrometer for spectral reflectance measurement, only spectral features were imported as predictors in our model. Thus, the accuracies of SCR detection and severity models could likely be further improved by the combined utilization of spectral and textural features of imaging spectroscopy. In addition, considering that a hyperspectral sensor is relatively expensive, a less cost-intensive multispectral sensor mounted on UAV platform could also be potentially used for SCR monitoring over large areas, but with lower accuracies [25,52,62]. We thus recommend further studies on the developed SDIs for large-scale remote sensing of SCR at different study sites and maize growing stages with UAVs-collected imaging spectroscopy or less complex multispectral datasets [52]. Another important direction of future studies to improve disease management is the early detection of crop diseases (e.g., SCR) before the symptoms become visible [6,8,63,64]. However, due to the limited data size of this study, we did not have a chance to explore the changes in the spectral signal of SCR over time compared to the earliest SCR detection time using our proposed method. In the future, a greenhouse-based experiment, which could provide a more control environment and more samples, could be combined with field measurements to explore this direction [64,65].

Conclusions
In this study, we analyzed the spectral signatures of healthy and SCR-infected corn leaves and developed spectral disease indices for detecting SCR occurrences and classifying SCR damage severity. To improve efficiency and accuracy, wavelengths (features), sensitive to SCR detection (572 nm, 766 nm, and 1445 nm), and severity classification (575 nm, 640 nm, and 1670 nm), were chosen for developing SDIs (i.e., HI and SI), respectively, using the RELIEF-F algorithm, instead of the full range of spectral measurement (i.e., VIS-NIR-SWIR). The selected wavelengths concentrated on spectral regions associated with leaf chlorophyll content, leaf internal structure, and leaf water content. Finally, using SVM and SDIs, as well as 38 stress-related VIs from literature, different SCR detection and SCR severity classification models were established. A 10-fold cross-validation strategy was applied to validate and compare the performances of different models in SCR detection and severity classification. The results showed that the SDIs-based models consistently had the highest accuracies among all models, in terms of both SCR detection (OA: 87%; Marco-F1 score: 0.856) and severity classification (OA: 70%; Marco-F1 score: 0.698). Our study thus highlights a fast and nondestructive method for SCR monitoring in the field. With the increasing use of UAVs with hyperspectral measurement capability, more studies should be conducted with our developed SDIs for remote sensing of SCR over large areas at different study sites and growing stages.
Supplementary Materials: The following are available online at http://www.mdpi.com/2072-4292/12/19/3233/s1. Table S1. 38 stress-related vegetation indices used in the study for SVM modeling; Table S2. Detailed information on leaf samples. Figure S1. Averaged confusion matrix of different SCR detection models using support vector machine based 10-fold cross validation: a. Health Index (HI); b. The ratio of first derivative values at 730-706 nm (D730/D706); c. Double Difference Index (DDI); d. The red-edge position through linear extrapolation (REP_LE); e. The ratio of first derivative values at 715-705 nm (D715/D705); f. MERIS Terrestrial Chlorophyll Index (MTCI); the fraction number in the matrix was from the 10-fold cross validation results divided by ten. Figure S2. Averaged confusion matrix of different SCR severity models using support vector machine based 10-fold cross validation: a. Severity Index (SI); b. Disease-Water Stress Index (DWSI); c. Physiological Reflectance Index (PRI); d. The simple ratio between the maxima of the first derivatives of reflectance at the red edge and green regions (EGFR); e. The normalized ratio between the maxima of the first derivatives of reflectance at the red edge and green regions (EGFN); f. Simple Ratio Index (SRI); the fraction number in the matrix was from the 10-fold cross validation results divided by ten. R source codes for SVM-based SCR detection and severity classification modeling; SCR samples in csv format.