The experiments were conducted at the University of Florida’s Southwest Florida Research and Education Center in Immokalee, FL, United States. Guidelines established by the University of Florida were followed for land preparation, fertility, irrigation, weed management, and insect control. The beds were 0.81 m wide with 3.66 m centers covered with black polyethylene film. Four-week-old watermelon transplants “Crimson Sweet” were planted on 9 March 2019 into the soil (Immokalee fine sand) in a complete randomized block treatment design with four replicates. Each plot consisted of 10 plants spaced 0.91 m apart within an 8.23 m row with 3.05 m between each plot. The plants were infected naturally by DM.

Healthy watermelon leaves were collected on April 10, 2019. After the first detection of DM, leaves were collected from DM-affected plants in five disease severities (DS; percentage of leaf tissue affected) stages. For the laboratory analysis, leaves were collected in the low DS stage (few lesions); then, from the medium 1 and 2, high and very high DS stages. The grading system of DM severity is as follows: the low DS had 5–10% severity, and as the disease progressed, the percentage of infection gradually increased in low to medium DS stages 1 and 2 (11–20% and 21–30% severity, respectively), and then the high and very high DS stages increased dramatically from 31 to 50% and 51 to 75% severity, respectively (Figure 1). In this study, spectral data were not collected when the DS was more than 75%, because the leaves were in very bad shape and desiccated. In the field (UAV-based) study, DS was categorized in two stages: low and high (Figure 2). The difference between the number of DS stages in the laboratory and field was due to the fact that, in the laboratory, each leaf was analyzed as a sample, and in the field, an entire plant was used as a sample. Hence, it was easier to quantify the DS stage in a single leaf than in an entire plant, which can include leaves in different DS stages.

Spectral data were collected using a benchtop hyperspectral imaging system, Pika L 2.4 (Resonon Inc., Bozeman MT, United States) (Figure 1E). The Pika L 2.4 was equipped with a 23-mm lens, which has a spectral range of 380–1,030 nm, 15.3° field of view, and a spectral resolution of 2.1 nm. The same hyperspectral camera was utilized in the laboratory (Figure 1E) and field (Figure 2C) after changing lenses, which covered the same spectral range. Resonon’s hyperspectral imagers (RHI), known as push-broom imagers, are line-scan imagers, which have 281 spectral channels. The system is made up of a linear stage assembly, which is shifted by a stage motor. In the laboratory, controlled broadband halogen lighting sources were set up above the linear stage to produce ideal situations for conducting spectral scans. The hyperspectral imaging system was arranged in a way that the lens’ distance from the linear stage was 0.5 m. The lights were positioned at the same level as the lens on a parallel plane. All scans were performed using the Spectronon Pro (Resonon Inc., Bozeman, MT, United States) software, which was connected to the camera system using a USB cable. Before performing the scans of the leaves, dark current noise was removed using the software. Then, the camera was calibrated by using a white tile (reflectance reference), provided by the manufacturer, and placed under the same conditions as used for performing scans. The selection of the regions of interest (RoIs) (e.g., Figure 1B) was done manually by picking six spectral scan regions per leaf (10 leaves per DS stage) to prevent the occurrence of any bias. The total spectral scans (RoIs) selected for each DS stage was 60. Regions of interest were selected in such a way that they included both the affected and unaffected areas of leaf tissue. The pixel number of each spectral scan selected was between 800 and 900 pixels. The average of 60 spectral scans was used to form an overall spectral scan signature curve for each DS stage. The Spectronon Pro software, which is a post-processing data analysis software, was used to analyze the spectral data of each leaf scan. Several areas containing the symptomatic and non-symptomatic regions on the leaves were selected using the selection tool and the spectrum was generated. For the healthy and DM-affected plans (five DS stages), several random spots on leaves were selected and the average spectral reflectance was calculated and used to form the spectral signature curves.

Spectral data were collected in the field by using a UAV (Matrice 600 Pro, Hexacopter, DJI Inc., Shenzhen, China) and the same hyperspectral camera (Pika L 2.4). The UAV-based imaging system included (Figure 2C): (i) a Resonon Pika L 2.4 hyperspectral camera; (ii) a visible-NIR (V-NIR) objective lens for the Pika L camera with a focal length of 17 mm, field of view (FOV) of 17.6 degrees, and an instantaneous field of view (IFOV) of 0.71 mrad; (iii) a Global Navigation Satellite System (GNSS) (Tallysman 33-2710NM-00-3000, Tallysman Wireless Inc., Ontario, Canada)/Inertial Measurement Unit (IMU) (Ellipse N, SGB Systems S.A.S., France) flight control system for multirotor aircraft to record sensor position and orientation, and (iv) a Resonon hyperspectral data analysis software (Spectronon Pro, Resonon, Bozeman, MT, United States), which is capable of rectifying the GPS/IMU data using a georectification plugin. The data were collected at 30 m above the ground and a speed rate of 1.5 m s^–1. In the produced map, the pixel size is a function of the working distance (distance between the camera lens and the scanning stage/field) and FOV. This value varies according to the flight parameters. In this study, it was around 35 mm per pixel. Gray tarp (Group VIII Technologies) was utilized to correct the data reflectance from radiance; the reflectance tarp was 36%. Radiometric calibration was performed by using a calibrated integrating sphere. The manufacturer took 100 lines of spectral data and built a radiometric calibration file that contains a lookup table with all combinations of integration times and frame rates. These data were used to convert raw camera data (digital numbers) to physical units of radiance in micro flicks. The Pika L 2.4 camera is a “pushbroom” or line-scan type imager that produces a 2-D image, where every pixel in the image contains a continuous reflectance spectrum. A calibration tarp was used to calibrate the data for various illumination conditions in the field (Figure 2D). The RoIs were randomly handpicked for each plant, and several spectral scans were done to cover the entire canopy. Each RoI contained four pixels, and four RoIs were selected for each plant. The total sample size for each DS stage was 20 plants. The RoIs were then transferred as a text file and processed using the SPSS software (SPSS 13.0, Inc., Chicago; Microsoft Corp., Redmond, WA, United States).

Vegetation indices could serve as indicators to identify DS stages based on any defectiveness in the functioning of plants such as physiochemical defects that affect photosynthesis, metabolic, and nutritional processes. The factors that are most affected by diseases and that could be measured are chlorophyll content, cell structure, cell sap, presence, and relative abundance of pigments concentration, water content, and carbon as expressed in the solar-reflected optical spectrum (400–2,500 nm) (Xiao et al., 2014). A neural network multilayer perceptron was performed to select the best VIs that could identify DM disease and its DS. For data analysis and evaluation of all VIs evaluated in this study (Table 1), the SPSS software was used. The correlation coefficient was another parameter that was utilized in this study to evaluate its VI’s performance in detecting DM severity stages.

During the spectral data collection under optimal light and temperature conditions in the laboratory, the spectral reflectance of five DM disease severity stages was taken. The spectral signatures and correlation coefficient at different disease severities were measured and compared (Figure 3). The spectral reflectance of the very high DS stage showed a significant increase in the green and red range (450–650 nm), while other stages showed lower spectral reflectance in the visible range (380–700 nm). The spectral reflectance values of the very high DS stage were decreased in the NIR range, and the spectral reflectance in the red edge diverged from other stages. The leaves of watermelon in the low and medium DS stages had very few symptoms; therefore, there were only slight differences between the spectral reflectance for these stages. The spectral reflectance signature is strongly related to the severity of the expected DM pathogen symptom (Figure 3A). In the visible range (380–700 nm), there were not many differences in the correlation coefficient signature of all DS stages (Figure 3B). The signatures were identical until the red-edge range (700 nm), where it gradually showed differences between DS stages (especially the high and very high severity stages). It is obvious from Figure 3B that the correlation coefficient showed wide differences in the NIR range as the DS increased.

The results of the classification varied based on disease severity, as in the low DS stage, the classification was lower than the other stages (e.g., 62.3 and 52.6% for MLP and DT, respectively). The classification rate increased as the severity of disease symptoms increased (Figure 4). The highest classification value of MLP was in the very high DS stage at 90%. The DT classification method had lower detection accuracies than the MLP classification method for all DS stages (Figure 4). Hence, the MLP method was selected as the best classification method for DM detection and DS stages classification. Therefore, it was used for the selection of the best wavebands and VIs for disease detection. The best wavebands for detecting the low and medium 1 DS stages were found on the red edge (711–722 nm), medium 2 and high DS stages were found at 1,007–1,020 nm, and the very high DS stage was found in red edge (759–761 nm). The best VIs, selected by using the MLP classification method, for low and medium DS stages 1 were the Cl green and PRI, respectively, while for medium 2, high, and very high DS stages were the NPQI (Table 2).

Figure 5A shows the UAV-based spectral signatures of healthy and DM-affected plants. In the visible range, the spectral reflectance values of healthy plants and DM-affected plans in a low DS stage were lower than in the high DS stage, which had a peak value of 20% at the green band. The spectral reflectance values of healthy plants were lower than the low and high DS stages in the NIR range. The spectral reflectance of the low DS stage was higher than the high DS stage, especially in the range of 750–915. Major differences cannot be seen in the red edge, where both DS stages had almost identical signatures. In the NIR range (700–1,000 nm), the spectral reflectance values of the low and high DS stages were higher than the spectral reflectance values of healthy plants.

The correlation coefficients for healthy, low, and high DS stages were almost identical in the visible range (Figure 5B). In the range of 750–1,000 nm, there were some differences recorded between the correlation coefficient values of the two DS stages (Figure 5B). Figure 5B shows differences between the low and high DS stages at 750 and 1,000 nm.

The MLP method gave a higher classification accuracy than the DT method; in the low DS stage, the classification accuracy of the MLP was 69%, while the classification accuracy of the DT was 60% (Figure 6). The highest classification accuracy was achieved in the high DS stage at 91% for MLP, while it was 69% for DT. The best wavebands for low DS stage classification were between 952 and 965 nm, while in the high DS stage the best wavebands were in red edge (755–780 nm). The best VIs for disease detection were the RARSc and RARSa for the low DS stage, and the RARSb and CI green for the high DS stage (Table 2).

The DM-affected watermelon leaves showed varied spectral reflectance signatures for each DS stage (both in the laboratory and field). However, in the visible range, there were no significant differences between the healthy and the low DS stage (5–10% infection), which indicates that it is very difficult to distinguish among these stages with visual observation. For that reason, hyperspectral imaging and machine learning can be used for a more efficient and rapid early plant disease and severity detection (compared to visual detection methods).

Laboratory measurements showed that the spectral reflectance of healthy plants was higher than the other DS stages in the blue range. The main differences between healthy and infected plants in the spectral signatures, both in the laboratory and field measurements, were found in the high severity stage in the green, red edge, and NIR range (700–1,000), while in NIR, the spectral reflectance signature of healthy plants was higher than the other DS stages. The high DS stage had a unique signature that can be used to distinguish this stage from others (and healthy plants). Although we were able to visually observe the change of the colors of the leaves for some of the DS stages, not all had the same level of change.

In the field, the results did not show significant differences between the low and high DS stages in the visible range; only slight differences were observed. However, that minor differences in the spectral reflectance in the visible range can be still considered as an indication of color change in the leaves. The leaves and the plant canopy in the low DS stage showed very few visual symptoms, and the classification accuracy was low, both in the laboratory and field. As the chlorophyll content and water content decrease, the leaf cell damage increases (Barnawal et al., 2017), and that helps the classification methods to better classify the DS stages, especially in the medium and very high DS stages.

The common practice for selecting significant wavelengths for the DS detection of new VIs is by the correlation to a biochemical or biophysical trait, for example, chlorophyll a + b content leaf structure parameter, the water content, and so on (Gitelson and Merzlyak, 1996; Hatfield et al., 2008). Regularly, the fluctuation of spectral reflectance might guide to the detection of plants under stress without specifying or providing a description of what type of stress cause the damage to the plant (Carter and Miller, 1994; Gitelson and Merzlyak, 1994). For example, as was mentioned earlier, several factors could reduce the chlorophyll content, of which some are physiological or biochemical caused by DM severity and this reduction might influence the photosynthesis activity.

In both experimental conditions, the best vegetation indices (e.g., the PRI, this index is more related to the green range; and the CI green) were able to discriminate between healthy and DM-affected plants in the low and the high DS stages. The PRI, which is produced by normalizing 531 and 570 nm, basically relays on the green range. Any deficiency or disorder in chlorophyll will influence the PRI value; the PRI has increasingly been used as an indicator of photosynthetic efficiency (Garbulsky et al., 2011) and as an indicator of water stress (Suarez et al., 2008). One of the first changes in a DM-affected plant is the reduction of the chlorophyll concentration that affects the process of photosynthesis in the infected leaf, and some VIs associated with chlorophyll content could be used to detect these changes (Mandal et al., 2009; Bellow et al., 2013; Barnawal et al., 2017). Our findings suggest the same.