*2.2. Data Collection and Preprocessing*

(**a**) (**b**) Data collection and preprocessing was followed from the method of Ashapure et al. (2019) [31]. UAS data (both MS and RGB) were collected over the experimental field on a weekly basis. Table 2 presents the flight specifications for both RGB and MS data collection. A total of eleven and ten flights were conducted in 2017 and 2018, respectively, using RGB and MS sensors. The overlap for MS sensor data collection was 70%, and it was 80%–85% for RGB sensor data collection. In 2017, the altitude

was about 20 m for the RGB sensor and 25 m for the MS sensor. In 2018, the flight altitude for RGB and MS sensors was 35 and 47 m, respectively. Provided the experimental field was in a coastal area, wind speed and rain were the potential factors to be considered before every flight. Most flights were conducted between 10:00AM and 2:00PM, except under unfavorable weather conditions, such as a wind speed greater than 15 mph or raining. Moreover, the temperature variation throughout the growing season varied between 79 and 96 ◦F.


**Table 2.** UAS data collection timeline and sensor-wise flight specification.

Generally, UAS are equipped with a consumer grade global positioning system (GPS) that do not have satisfactory location accuracy for aerial mapping applications. To overcome this problem, semi-permanent ground control points (GCPs) with a high reflectance were installed over the study area. The GCPs were surveyed using a dual frequency, post processed kinematic (PPK) GPS system, 20 Hz V-Map Air model (Micro Aerial Project L.L.C., Gainesville, FL). Images obtained from the UAS platform with significant overlaps along with the 3D coordinates of the GCPs were imported to Agisoft Photoscan Pro (Agisoft LLC, St. Petersburg, Russia), which uses structure from motion (SfM) photogrammetry algorithms to derive high dens 3D point clouds, fine spatial resolution 2D orthomosaics, and digital surface models (DSM). The SfM refers to the process of finding the three-dimensional structure of an object by analyzing local motion signals over time [43].

#### *2.3. Canopy Cover Computation*

The percentage CC was computed as the ratio of canopy area to the total area of the grid computed using Equation (1), where GSD is ground sampling distance. An RGB orthomosaic image was converted into a binary image, where zero represents non-canopy pixels and one represents canopy pixels. As the whole field was divided into a square meter grid by using Equation (1), a grid-wise percentage CC was computed (As shown in Figure 3).

$$\text{ACC} = \left(\frac{\sum \text{(GCD}^2\text{)} \text{ if } \text{Cauchy } \text{Pixel}}{\sum \text{(GCD}^2\text{)}}\right) \times 100,\tag{1}$$

wise percentage CC was computed (As shown in Figure 3).

CC =൬

∑൫൯

The percentage CC was computed as the ratio of canopy area to the total area of the grid computed using Equation (1), where GSD is ground sampling distance. An RGB orthomosaic image was converted into a binary image, where zero represents non-canopy pixels and one represents canopy pixels. As the whole field was divided into a square meter grid by using Equation (1), a grid-

**Figure 3.** Canopy cover estimation from the orthomosaic images (Red square: individual crop grid, and each grid is 1 × 1 m): an RGB orthomosaic image collected using the unmanned aerial systems (UAS) platform, followed by the binary classification results of the orthomosaic image, where white represents the canopy class and black represents the non-canopy class; the last image represents the grid-wise estimated canopy cover (CC). **Figure 3.** Canopy cover estimation from the orthomosaic images (Red square: individual crop grid, and each grid is 1 × 1 m): an RGB orthomosaic image collected using the unmanned aerial systems (UAS) platform, followed by the binary classification results of the orthomosaic image, where white represents the canopy class and black represents the non-canopy class; the last image represents the grid-wise estimated canopy cover (CC).

As mentioned earlier, MS sensor-based CC estimation is considered, in the literature, as the most reliable estimation technique that uses a normalized difference vegetation index (NDVI) to separate the canopy from the non-canopy areas (computed using Equation (2) [44]). SlantView, the software developed for the SlantRange 3p MS sensor, was used for the radiometric calibration in order to accurately compare the crop conditions across datasets collected in varying lighting conditions throughout the day and growing season. A detailed visual inspection was performed to find a threshold NDVI value to separate the canopy area from the non-canopy area in the image throughout the growing season regardless of the growth stage. As mentioned earlier, MS sensor-based CC estimation is considered, in the literature, as the most reliable estimation technique that uses a normalized difference vegetation index (NDVI) to separate the canopy from the non-canopy areas (computed using Equation (2) [44]). SlantView, the software developed for the SlantRange 3p MS sensor, was used for the radiometric calibration in order to accurately compare the crop conditions across datasets collected in varying lighting conditions throughout the day and growing season. A detailed visual inspection was performed to find a threshold NDVI value to separate the canopy area from the non-canopy area in the image throughout the growing season regardless of the growth stage.

$$\text{NDVI} \, = \left( \frac{\text{NIR} - \text{Red}}{\text{NIR} + \text{Red}} \right) \, \tag{2}$$

∑൫൯ ൰ × 100, (1)

To investigate the accurate CC estimation using the RGB-based sensor, which is equivalent to CC estimation using NDVI, a pixel-wise classification method was implemented; this is presented in Figure 4. As found in the literature, pixel classification methods are considered highly accurate for separating the canopy and non-canopy classes, and they are mainly used to calibrate RGB-based methods [35,38,45]. A pixel classification method based on K-means clustering was used to compare the RGB-based methods that use vegetation indices to separate canopy areas from non-canopy areas. Initially, K-means clustering with five classes was applied to the RGB orthomosaics, considering five potential classes representing soil, shadow, cotton bolls, green canopy and brown canopy, as presented in Figure 4. After assigning the class labels to the clustered map, it was validated using ground-truth samples collected over RGB orthomosaics using visual inspection, and the overall classification accuracy was found to be at least 97%. Later, soil, shadow and cotton bolls were merged and assigned as non-canopy, and green and brown canopy were merged and identified as canopy classes. However, as pixel-wise classification-based CC estimation is computationally extensive, its implementation was solely done with an intention to investigate the maximum achievable performance using the RGB-based sensors and to compare it with MS sensor-based CC estimation. To investigate the accurate CC estimation using the RGB-based sensor, which is equivalent to CC estimation using NDVI, a pixel-wise classification method was implemented; this is presented in Figure 4. As found in the literature, pixel classification methods are considered highly accurate for separating the canopy and non-canopy classes, and they are mainly used to calibrate RGB-based methods [35,38,45]. A pixel classification method based on K-means clustering was used to compare the RGB-based methods that use vegetation indices to separate canopy areas from non-canopy areas. Initially, K-means clustering with five classes was applied to the RGB orthomosaics, considering five potential classes representing soil, shadow, cotton bolls, green canopy and brown canopy, as presented in Figure 4. After assigning the class labels to the clustered map, it was validated using ground-truth samples collected over RGB orthomosaics using visual inspection, and the overall classification accuracy was found to be at least 97%. Later, soil, shadow and cotton bolls were merged and assigned as non-canopy, and green and brown canopy were merged and identified as canopy classes. However, as pixel-wise classification-based CC estimation is computationally extensive, its implementation was solely done with an intention to investigate the maximum achievable performance using the RGB-based sensors and to compare it with MS sensor-based CC estimation.

*Remote Sens.* **2019**, *10*, x FOR PEER REVIEW 7 of 17

**Figure 4.** K-means clustering-based pixel classification method workflow where the orthomosaic is classified into five classes, and, later, classes are merged into two clusters, namely canopy and noncanopy. The RGB orthomosaic presented was captured on 19th June 2017. **Figure 4.** K-means clustering-based pixel classification method workflow where the orthomosaic is classified into five classes, and, later, classes are merged into two clusters, namely canopy and non-canopy. The RGB orthomosaic presented was captured on 19th June 2017.

In this study, four different RGB-based methods were used, namely Canopeo, the excessive greenness index (ExG), the modified green red vegetation index (MGRVI) and the red green blue vegetation index (RGBVI), to generate the binary images to separate canopy areas from non-canopy areas (Table 3). In this study, four different RGB-based methods were used, namely Canopeo, the excessive greenness index (ExG), the modified green red vegetation index (MGRVI) and the red green blue vegetation index (RGBVI), to generate the binary images to separate canopy areas from non-canopy areas (Table 3).


**Table 3.** RGB image-based vegetation indices and their formula.

MGRVI <sup>ଶ</sup> − <sup>ଶ</sup> ଶ + ଶ [47] RGBVI <sup>ଶ</sup> −× ଶ +× [47] The overall procedure to compute the CC binary map using RGB vegetation indices is presented in Figure 5. The Canopeo algorithm resulted in a binary map that separated canopy areas from noncanopy areas; however, applying vegetation indices over the RGB mosaics resulted in a grayscale image. Similar to the NDVI, an empirical evaluation was performed over all the grey scale vegetation index maps to decide a threshold value for the ExG, the MGRVI and the RGBVI that could separate canopy areas from non-canopy areas. A detailed visual inspection was performed to determine a The overall procedure to compute the CC binary map using RGB vegetation indices is presented in Figure 5. The Canopeo algorithm resulted in a binary map that separated canopy areas from non-canopy areas; however, applying vegetation indices over the RGB mosaics resulted in a grayscale image. Similar to the NDVI, an empirical evaluation was performed over all the grey scale vegetation index maps to decide a threshold value for the ExG, the MGRVI and the RGBVI that could separate canopy areas from non-canopy areas. A detailed visual inspection was performed to determine a threshold value to separate canopy areas from non-canopy areas in the image throughout the growing season, regardless of the growth stage. Considering the homogeneity of the crop, only a subset of the image was used to determine the threshold, and the threshold chosen for each VI is presented in Table 4. The demonstration of visual inspection is presented in Figure 6, where a subset of an early

stage and a mature stage RGB image of the same area is considered. Originally for all the RGB images in the growing season, a range of threshold values with a step size of 0.01 was applied to a subset area in the images to generate binary map using the grayscale VI map of the subset area. However, for the demonstration, only one VI (ExG) image was considered, and this was generated using one early stage (image taken on 7 June 2017) and one mature stage (image taken on 10 July 2017) RGB image with a range of threshold values and a step size of 0.02. It can be observed from Figure 6 that variation in threshold values did not affect the binarization of the early stage image, as most of the canopy was green. However with the higher threshold (0.22), the binary image had some canopy pixels not classified as canopy due to their darker color. The effect of variation in threshold was more significant in the mature stage image. A lower threshold value of 0.18 resulted in a lot of non-canopy pixels classified as canopy, especially the shadow pixels. A slightly higher threshold value of 0.22 resulted in more conservative classification that omitted a substantial number of canopy pixels. Visual inspection suggested that the threshold value of 0.2 resulted in a most appropriate classification for the ExG. Similarly, other VIs were also examined under visual inspection to select a single threshold value for all the images in the growing season. Figure 6 that variation in threshold values did not affect the binarization of the early stage image, as most of the canopy was green. However with the higher threshold (0.22), the binary image had some canopy pixels not classified as canopy due to their darker color. The effect of variation in threshold was more significant in the mature stage image. A lower threshold value of 0.18 resulted in a lot of non-canopy pixels classified as canopy, especially the shadow pixels. A slightly higher threshold value of 0.22 resulted in more conservative classification that omitted a substantial number of canopy pixels. Visual inspection suggested that the threshold value of 0.2 resulted in a most appropriate classification for the ExG. Similarly, other VIs were also examined under visual inspection to select a single threshold value for all the images in the growing season. **Table 4.** Threshold chosen for vegetation indices (Vis) to separate canopy and non-canopy areas. **VI Threshold Range**  NDVI 0.6 0 to 1 ExG 0.2 −2 to 2 MGRVI 0.15 −1 to 1 RGBVI 0.15 −1 to 1

*Remote Sens.* **2019**, *10*, x FOR PEER REVIEW 8 of 17

growing season, regardless of the growth stage. Considering the homogeneity of the crop, only a subset of the image was used to determine the threshold, and the threshold chosen for each VI is presented in Table 4. The demonstration of visual inspection is presented in Figure 6, where a subset of an early stage and a mature stage RGB image of the same area is considered. Originally for all the RGB images in the growing season, a range of threshold values with a step size of 0.01 was applied to a subset area in the images to generate binary map using the grayscale VI map of the subset area. However, for the demonstration, only one VI (ExG) image was considered, and this was generated

2017) RGB image with a range of threshold values and a step size of 0.02. It can be observed from

**Figure 5.** Procedure to generate a binary map that indicates canopy and non-canopy areas. Applying the vegetation index over the RGB orthomosaic resulted in a grayscale image. By applying thresholding, a binary image was generated. Lastly, morphological closing was applied over binary image to improve binary classification. The presented RGB orthomosaic was captured on 10 July 2017, and the excessive greenness index (ExG) was the VI used for the demonstration of the methodology. **Figure 5.** Procedure to generate a binary map that indicates canopy and non-canopy areas. Applying the vegetation index over the RGB orthomosaic resulted in a grayscale image. By applying thresholding, a binary image was generated. Lastly, morphological closing was applied over binary image to improve binary classification. The presented RGB orthomosaic was captured on 10 July 2017, and the excessive greenness index (ExG) was the VI used for the demonstration of the methodology.

**Table 4.** Threshold chosen for vegetation indices (Vis) to separate canopy and non-canopy areas.


RGB-based vegetation indices accurately identified healthy green canopy; however, later in the season as the canopy started to change the color, their ability to identify canopy deteriorated. To further improve the binary map (indicating canopy and non-canopy areas), a morphological closing operation

was performed. The morphological closing operation is a combination of dilation and erosion, and it helps to remove small holes while keeping the separation boundary intact [48]. For this experiment, a 3 × 3 kernel window over one iteration was used to perform closing operation. *Remote Sens.* **2019**, *10*, x FOR PEER REVIEW 9 of 17

**Figure 6.** Procedure to select appropriate threshold value to generate a binary map that indicates canopy and non-canopy areas. The first images are a subset of RGB images captured on 7 June 2017 and 10 July 2017. Second images are the result of applying the vegetation index (ExG) over RGB images. The next three images are the result of applying varying threshold values that were superimposed on the original RGB images (the canopy classified pixels are represented by the red color, and non-canopy pixels were set to transparent). **Figure 6.** Procedure to select appropriate threshold value to generate a binary map that indicates canopy and non-canopy areas. The first images are a subset of RGB images captured on 7 June 2017 and 10 July 2017. Second images are the result of applying the vegetation index (ExG) over RGB images. The next three images are the result of applying varying threshold values that were superimposed on the original RGB images (the canopy classified pixels are represented by the red color, and non-canopy pixels were set to transparent).

#### **3. Results**

RGB-based vegetation indices accurately identified healthy green canopy; however, later in the season as the canopy started to change the color, their ability to identify canopy deteriorated. To further improve the binary map (indicating canopy and non-canopy areas), a morphological closing operation was performed. The morphological closing operation is a combination of dilation and erosion, and it helps to remove small holes while keeping the separation boundary intact [48]. For this experiment, a 3 × 3 kernel window over one iteration was used to perform closing operation. **3. Results**  The CC grid maps at each flight for both 2017 and 2018 are presented in Figures 7 and 8, respectively. The visual inspection of the grid maps revealed that the percentage canopy cover increased in the growing season and reached its plateau right after the middle of the season (19 June for the 2017 experiment and 13 June for the 2018 experiment). Later in the season, it was observed that the percentage CC started to slightly decrease with the canopy senescence. For the 2017 experiment, a rapid decay in CC was observed between 18 July and 23 July due to a common practice in the cotton fields known as defoliation, which prepares the crop for harvesting. A similar effect was observed in the 2018 experiment between 9 July and 19 July.

The CC grid maps at each flight for both 2017 and 2018 are presented in Figures 7 and 8, respectively. The visual inspection of the grid maps revealed that the percentage canopy cover increased in the growing season and reached its plateau right after the middle of the season (19 June for the 2017 experiment and 13 June for the 2018 experiment). Later in the season, it was observed that the percentage CC started to slightly decrease with the canopy senescence. For the 2017 experiment, a rapid decay in CC was observed between 18 July and 23 July due to a common practice in the cotton fields known as defoliation, which prepares the crop for harvesting. A similar effect was observed in the 2018 experiment between 9 July and 19 July. Following the methodology presented in Figure 4 using RGB images, K-means clustering-based classification maps were generated considering five clusters which were later labeled to represent soil, cotton boll, shadow, green canopy, and brown canopy. Later, binary maps were generated by merging soil, cotton boll and shadow classes to indicate non-canopy, while brown canopy and green canopy classes were merged to indicate canopy pixels. The comparison of the average CC per grid using the NDVI and K-means (also referred as RGB reference) is presented in Figures 9 and 10 for the Following the methodology presented in Figure 4 using RGB images, K-means clustering-based classification maps were generated considering five clusters which were later labeled to represent soil, cotton boll, shadow, green canopy, and brown canopy. Later, binary maps were generated by merging soil, cotton boll and shadow classes to indicate non-canopy, while brown canopy and green canopy classes were merged to indicate canopy pixels. The comparison of the average CC per grid using the NDVI and K-means (also referred as RGB reference) is presented in Figures 9 and 10 for the 2017 and 2018 experiments, respectively. From both the 2017 and 2018 experiments, it was observed that the NDVI-based and RGB reference-based average CC per grid followed the same trend throughout the growing season, and there was a one-to-one correspondence between the two when plotted as a straight line at an intercept of one with very high R<sup>2</sup> values (0.98 for 2017 and 0.97 for 2018). The results indicated that it is possible to achieve the same level of performance using the RGB-based sensor as that of the MS sensors using NDVI. However, the purpose of the RGB reference CC estimation was only to provide a comparison reference for the MS sensor-based CC estimation.

estimation was only to provide a comparison reference for the MS sensor-based CC estimation.

2017 and 2018 experiments, respectively. From both the 2017 and 2018 experiments, it was observed that the NDVI-based and RGB reference-based average CC per grid followed the same trend throughout the growing season, and there was a one-to-one correspondence between the two when

dataset.

dataset.

*Remote Sens.* **2019**, *10*, x FOR PEER REVIEW 10 of 17

**Figure 7.** CC grid maps generated at each flight in growing season using normalized difference vegetation index (NDVI) maps for the 2017 dataset. **Figure 7.** CC grid maps generated at each flight in growing season using normalized difference vegetation index (NDVI) maps for the 2017 dataset. **Figure 7.** CC grid maps generated at each flight in growing season using normalized difference vegetation index (NDVI) maps for the 2017 dataset.

**Figure 8.** CC grid maps generated at each flight in growing season using NDVI maps for the 2018 **Figure 8.** CC grid maps generated at each flight in growing season using NDVI maps for the 2018 dataset.

*Remote Sens.* **2019**, *10*, x FOR PEER REVIEW 11 of 17

*Remote Sens.* **2019**, *10*, x FOR PEER REVIEW 11 of 17

**Figure 9.** For the 2017 experiment: (**a**) the average NDVI and RGB reference-based percentage CC for each flight in the growing season; (**b**) a comparison of the NDVI and RGB reference-based percentage CC techniques with R2. **Figure 9.** For the 2017 experiment: (**a**) the average NDVI and RGB reference-based percentage CC for each flight in the growing season; (**b**) a comparison of the NDVI and RGB reference-based percentage CC techniques with R<sup>2</sup> . **Figure 9.** For the 2017 experiment: (**a**) the average NDVI and RGB reference-based percentage CC for each flight in the growing season; (**b**) a comparison of the NDVI and RGB reference-based percentage CC techniques with R2.

**Figure 10.** For the 2018 experiment: (**a**) the average NDVI and RGB reference-based percentage CC for each flight in the growing season; (**b**) a comparison of the NDVI and RGB reference-based percentage CC techniques with R2. **Figure 10.** For the 2018 experiment: (**a**) the average NDVI and RGB reference-based percentage CC for each flight in the growing season; (**b**) a comparison of the NDVI and RGB reference-based percentage CC techniques with R2. **Figure 10.** For the 2018 experiment: (**a**) the average NDVI and RGB reference-based percentage CC for each flight in the growing season; (**b**) a comparison of the NDVI and RGB reference-based percentage CC techniques with R<sup>2</sup> .

Using RGB images, four thresholding-based CC estimation methods were implemented, namely Canopeo, the ExG, the MGRVI and the RGBVI. Along with the NDVI-based average CC estimation per grid, the average CC estimation per grid in the growing season for each RGB-based CC estimation method before and after applying morphological closing (MC) is presented in Figures 11 and 12 for the 2017 and 2018 experiments, respectively. It can be observed from Figure 11 that, early in the growing season, the average CC estimated by all the methods was in agreement with very less variation and followed the same increasing trend as the NDVI-based estimation. However, the Canopeo and ExG methods reached their peak early in the season (7th June), as compared to other methods. The main reason for this was that they were accurately identifying healthy green canopy, but later, as the canopy started to change color from green to yellow and eventually to brown, they were not as efficient as other methods to identify the matured canopy. Moreover, after they reached their peak, they rapidly started to decrease later in the season, commensurate with the rate of change Using RGB images, four thresholding-based CC estimation methods were implemented, namely Canopeo, the ExG, the MGRVI and the RGBVI. Along with the NDVI-based average CC estimation per grid, the average CC estimation per grid in the growing season for each RGB-based CC estimation method before and after applying morphological closing (MC) is presented in Figures 11 and 12 for the 2017 and 2018 experiments, respectively. It can be observed from Figure 11 that, early in the growing season, the average CC estimated by all the methods was in agreement with very less variation and followed the same increasing trend as the NDVI-based estimation. However, the Canopeo and ExG methods reached their peak early in the season (7th June), as compared to other methods. The main reason for this was that they were accurately identifying healthy green canopy, but later, as the canopy started to change color from green to yellow and eventually to brown, they were not as efficient as other methods to identify the matured canopy. Moreover, after they reached their peak, they rapidly started to decrease later in the season, commensurate with the rate of change Using RGB images, four thresholding-based CC estimation methods were implemented, namely Canopeo, the ExG, the MGRVI and the RGBVI. Along with the NDVI-based average CC estimation per grid, the average CC estimation per grid in the growing season for each RGB-based CC estimation method before and after applying morphological closing (MC) is presented in Figures 11 and 12 for the 2017 and 2018 experiments, respectively. It can be observed from Figure 11 that, early in the growing season, the average CC estimated by all the methods was in agreement with very less variation and followed the same increasing trend as the NDVI-based estimation. However, the Canopeo and ExG methods reached their peak early in the season (7th June), as compared to other methods. The main reason for this was that they were accurately identifying healthy green canopy, but later, as the canopy started to change color from green to yellow and eventually to brown, they were not as efficient as other methods to identify the matured canopy. Moreover, after they reached their peak, they rapidly started to decrease later in the season, commensurate with the rate of change of color in the canopy in the later

of color in the canopy in the later season. As can be observed from Table 5, the average root mean

of color in the canopy in the later season. As can be observed from Table 5, the average root mean

season. As can be observed from Table 5, the average root mean square error (RMSE) of the percentage CC turned out to be highest amongst all (17.87 for Canopeo and 16.97 for the ExG). The MGRVI showed a slightly better performance over Canopeo and the ExG, and it was able to identify mature canopy. However, in comparison to the NDVI, it also reached its peak relatively early. The RGBVI turned out to be the most efficient method to estimate CC, as, especially later in the season, it outperformed the other RGB-based methods. However, the RGBVI still could not match the NDVI-based CC estimation. It was noticed that morphological closing significantly improved the CC estimation, and the average RMSE with the NDVI-based CC estimation was substantially reduced (Table 5). *Remote Sens.* **2019**, *10*, x FOR PEER REVIEW 12 of 17 and 16.97 for the ExG). The MGRVI showed a slightly better performance over Canopeo and the ExG, and it was able to identify mature canopy. However, in comparison to the NDVI, it also reached its peak relatively early. The RGBVI turned out to be the most efficient method to estimate CC, as, especially later in the season, it outperformed the other RGB-based methods. However, the RGBVI still could not match the NDVI-based CC estimation. It was noticed that morphological closing significantly improved the CC estimation, and the average RMSE with the NDVI-based CC

**Figure 11.** For the 2017 experiment, the average CC estimation per grid using the NDVI-based CC estimation throughout the growing season, along with the average CC estimation using (**a**) Canopeo, (**b**) the ExG, (**c**) the modified green red vegetation index (MGRVI) and (**d**) the red green blue vegetation index (RGBVI), before and after applying the morphological closing (MC) operation. **Figure 11.** For the 2017 experiment, the average CC estimation per grid using the NDVI-based CC estimation throughout the growing season, along with the average CC estimation using (**a**) Canopeo,(**b**) the ExG, (**c**) the modified green red vegetation index (MGRVI) and (**d**) the red green blue vegetation index (RGBVI), before and after applying the morphological closing (MC) operation.

(**a**) (**b**)

**4. Discussion** 

and 16.97 for the ExG). The MGRVI showed a slightly better performance over Canopeo and the ExG, and it was able to identify mature canopy. However, in comparison to the NDVI, it also reached its peak relatively early. The RGBVI turned out to be the most efficient method to estimate CC, as, especially later in the season, it outperformed the other RGB-based methods. However, the RGBVI still could not match the NDVI-based CC estimation. It was noticed that morphological closing significantly improved the CC estimation, and the average RMSE with the NDVI-based CC

(**a**) (**b**)

(**c**) (**d**) **Figure 11.** For the 2017 experiment, the average CC estimation per grid using the NDVI-based CC estimation throughout the growing season, along with the average CC estimation using (**a**) Canopeo,

vegetation index (RGBVI), before and after applying the morphological closing (MC) operation.

estimation was substantially reduced (Table 5).

**Figure 12.** For the 2018 experiment, the average CC estimation per grid using the NDVI-based CC estimation throughout the growing season, along with the average CC estimation using (**a**) Canopeo, (**b**) the ExG, (**c**) the MGRVI and (**d**) the RGBVI, before and after applying the morphological closing (MC) operation. **Figure 12.** For the 2018 experiment, the average CC estimation per grid using the NDVI-based CC estimation throughout the growing season, along with the average CC estimation using (**a**) Canopeo, (**b**) the ExG, (**c**) the MGRVI and (**d**) the RGBVI, before and after applying the morphological closing (MC) operation.



A similar trend was observed in the performance of thresholding-based CC estimation methods in the 2018 experiment (Figure 12). Canopeo and the ExG method had the higher RMSE, as compared to other methods (Table 5). However, CC estimation values of 2018 experiment were slightly higher compared to CC estimation values of 2017 experiment. The RGBVI turned out to be the most efficient CC estimation method amongst all. The application of the morphological closing operation significantly reduced the RMSE. It was observed that morphological closing operation resulted in slightly better performance in the 2018 experiment as compared to the 2017 experiment. It was A similar trend was observed in the performance of thresholding-based CC estimation methods in the 2018 experiment (Figure 12). Canopeo and the ExG method had the higher RMSE, as compared to other methods (Table 5). However, CC estimation values of 2018 experiment were slightly higher compared to CC estimation values of 2017 experiment. The RGBVI turned out to be the most efficient CC estimation method amongst all. The application of the morphological closing operation significantly reduced the RMSE. It was observed that morphological closing operation resulted in slightly better

filtering performance because a very high resolution leads to more details being observed, which could be undesirable when the information class under consideration is more general. Furthermore, a little difference in the CC growth pattern in two years was observed, which was a function of weather conditions (such as daily temperature and precipitation) and the type of genotype planted.

Recent years have witnessed an upsurge in UAS and sensor technology, an upsurge which has made it possible to collect high temporal and spatial resolution data over crops throughout the growing season. The main objective of this study was to provide a comparison framework between MS sensor-based CC estimation and RGB-based CC estimation, as very scarce attention has been paid to explore different VIs generated using UAS-based sensors to compute canopy cover in the literature. As mentioned in the literature, MS sensor-based CC estimation is more accurate and stable because it accounts for the live canopy based on the chlorophyll content present in the canopy, and performance in the 2018 experiment as compared to the 2017 experiment. It was noticed that in the 2018 experiments, the spatial resolution of the RGB images was slightly lower than the 2017 experiments. However, the difference in the spatial resolution was not much (difference of ~0.2cm), and a slight decrease of the GSD might contribute to the better classification and later filtering performance because a very high resolution leads to more details being observed, which could be undesirable when the information class under consideration is more general. Furthermore, a little difference in the CC growth pattern in two years was observed, which was a function of weather conditions (such as daily temperature and precipitation) and the type of genotype planted.

#### **4. Discussion**

Recent years have witnessed an upsurge in UAS and sensor technology, an upsurge which has made it possible to collect high temporal and spatial resolution data over crops throughout the growing season. The main objective of this study was to provide a comparison framework between MS sensor-based CC estimation and RGB-based CC estimation, as very scarce attention has been paid to explore different VIs generated using UAS-based sensors to compute canopy cover in the literature. As mentioned in the literature, MS sensor-based CC estimation is more accurate and stable because it accounts for the live canopy based on the chlorophyll content present in the canopy, and this content is highly reflected in the near-infrared (NIR) band; hence, the canopy varieties which are not green in color can be correctly accounted for. Moreover, changes in the color as the season progresses can also be identified accurately. However, the accuracy of the NDVI is a function of the type and quality of the multispectral sensor used to collect the image. In this study, temporal NDVI maps were comparable, despite changes in lighting conditions over different flights throughout the season, as they were generated from multispectral data collected using the SlantRange 3p sensor that performed radiometric calibration. The RGB-based CC estimation performed inadequately if the plant color deviated from green, which was confirmed by the experiments performed in both 2017 and 2018.

Except for the Canopeo method, all the other RGB VIs considered in this study required a threshold to be applied in order to separate canopy areas from non-canopy areas. In previous studies, the Otsu method has mostly been used for automatic thresholding [41]. The Otsu method resulted in an accurate thresholding early in the season, as the image had a mostly bimodal histogram and the variances of the spectral clusters were small compared to the mean difference. However, later in the season, thresholding by the Otsu method was questionable, because, as the season progressed, the variance in the spectral signature of the canopy increased, and, closer to senescence, the image no longer possessed a bimodal histogram due to the emergence of new spectral classes such as open cotton bolls. Consequently, in this study, the VIs were examined under visual inspection to select a single threshold value for all the images in the growing season. It was observed that the selected single threshold value successfully classified canopy pixels. Though they were less affected due to senescence as compared to the RGB-based VIs, the overall NDVI values still decreased in the late season; however, the selected single threshold value successfully classified canopy pixels with reasonable accuracy based on visual interpretation. Since this study was limited to the cotton crop, the threshold value might differ for other crops, and there might be a different trend observed in response to the senescence in the growing season. In future, an efficient thresholding method that can classify canopy regardless of growth stage would help automate the process.

As previous studies have affirmed that RGB-based CC estimation efficiently works early in the season [40,41], it was also observed in this study that early in the season, both the MS- and RGB-based CC estimations were in agreement and followed a similar increasing trend. Moreover, in previous studies, MS sensor-based CC estimation has been found to be more accurate in the later season as compared to RGB-based CC estimation [41]. In this study, it was observed that as the season progressed, MS sensor-based CC estimation kept on increasing, but RGB-based CC estimation peaked early and started to drop relatively rapidly for both the 2017 and 2018 experiments. Nevertheless, the RGBVI outperformed other RBG-based CC estimation methods, though still not close enough to

match the estimation by MS sensor-based CC estimation. That led to the question "is it possible to achieve the same level of accuracy by using RGB-based CC estimation as that of the NDVI-based CC estimation?" With the aforementioned objective, a K-means clustering-based CC estimation method was implemented, and this was tested by using ground truth samples for the canopy and non-canopy classes. It was observed that K-means clustering-based approach matched the accuracy level of the MS sensor-based CC estimation. However, there was a requirement to investigate any scope for an improvement to the RGB-based CC estimation approach, as the K-means clustering-based approach, or any classification-based approach is accurate but computationally extensive, especially in its parameter tuning and demand for ground truth sample collection, which is labor-intensive and time consuming. The objective of this study was to improve the RGB-based CC estimation approach. It was noticed that, later in the season, RGB-based indices were not able to capture the canopy pixels that were not green, which resulted in CC maps with a lot of small holes. The morphological closing operation proved to be a solution to this problem and helped to fill very small holes and keep the boundary of the canopy intact. In both the 2017 and 2018 experiments, it was noticed that applying the morphological closing operation significantly improved the performance of the RGB-based CC estimation. The RGBVI with morphological closing applied was found to have a CC estimation very close to the NDVI-based CC estimation. With the proposed approach, a more affordable alternate to the MS sensor can be provided to estimate CC.

### **5. Conclusions**

A comparative study was performed to evaluate CC estimation using an RGB sensor. With a multi-year CC analysis, MS sensor-based CC estimation was used as a reference, as it is considered a stable and accurate form of estimation. The correlation of RGB reference-based CC estimation with MS sensor-based CC estimation ensured the feasibility of using an RGB sensor to match the MS sensor-based CC estimation. An analysis of RGB-based methods suggested that the RGBVI was more tolerant to changes in the color of the canopy when the canopy started to senescence. Moreover, when applied with the morphological closing operation, the RGBVI-based CC estimation was found to be as accurate as MS sensor-based CC estimation, with an average RMSE of less than three percent. CC is a good predictor variable for plant growth parameters. As multispectral sensors are more sensitive and expensive, the proposed RGB-based CC estimation could provide an affordable alternate to agriculture scientists and breeders. In the future, this methodology will be investigated on other crops, as there might be a different trend observed in response to senescence in the growing season.

**Author Contributions:** conceptualization, A.A. and J.J.; data curation, A.A., A.C. and S.O.; formal analysis, A.A.; investigation, M.M.; methodology, A.A. and J.J.; project administration, J.J., M.M. and J.L.; resources, J.L.; software, A.A.; supervision, J.J.; validation, A.A.; visualization, A.A.; writing—original draft, A.A.; writing—review & editing, J.J., A.C., S.O. and J.L.

**Funding:** This research received no external funding.

**Acknowledgments:** This research was supported by Texas A&M AgriLife Research, Corpus Christi.

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

#### **References**


© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
