A Novel Index to Detect Vegetation in Urban Areas Using UAV-Based Multispectral Images

: Unmanned aerial vehicles (UAVs) equipped with high-resolution multispectral cameras have increasingly been used in urban planning, landscape management, and environmental monitoring as an important complement to traditional satellite remote sensing systems. Interest in urban regeneration projects is on the rise in Korea, and the results of UAV-based urban vegetation analysis are in the spotlight as important data to effectively promote urban regeneration projects. Vegetation indices have been used to obtain vegetation information in a wide area using the multispectral bands of satellites. UAV images have recently been used to obtain vegetation information in a more rapid and precise manner. In this study, multispectral images were acquired using a UAV equipped with a Micasense RedEde MX camera to analyze vegetation indices, such as the Normalized Difference Vegetation Index (NDVI), Green Normalized Difference Vegetation Index (GNDVI), Blue Normalized Difference Vegetation Index (BNDVI), Red Green Blue Vegetation Index (RGBVI), Green Red Vegetation Index (GRVI), and Soil Adjusted Vegetation Index (SAVI). However, in the process of analyzing urban vegetation using the existing vegetation indices, it became clear that the vegetation index values of long-run steel rooﬁng, waterproof coated roofs, and urethane-coated areas are often similar to, or slightly higher than, those of grass. In order to improve the problem of misclassiﬁcation of vegetation, various equations were tested by combining multispectral bands. Kappa coefﬁcient analysis showed that the squared Red-Blue NDVI index produced the best results when analyzing vegetation reﬂecting urban land cover. The novel vegetation index developed in this study will be very useful for effective analysis of vegetation in urban areas with various types of land cover, such as long-run steel rooﬁng, waterproof coated roofs, and urethane-coated areas.


Introduction
Globally, extreme heat events have been increasing in frequency and severity, resulting in negative impacts on human health [1][2][3]. At the regional scale, the "Urban Heat Islands" (UHIs) effect results in urban areas being hotter than nearby rural areas due to impervious surface cover, decreased cooling due to deforestation, anthropogenic heat release, and high concentrations of air pollutants, which can greatly impair air quality [4,5]. With the recent acceleration of urbanization, impervious surfaces, such as buildings, roads, and pavements, have come to completely dominate the urban space, causing an increase in surface temperature compared to that in surrounding rural areas due to their capacity to absorb and retain heat [6,7].
Vegetation considerably regulates local air temperatures [8,9]. However, nearly 88% of global primary vegetation-covered land in urban areas has been destroyed and replaced by artificial surfaces over the past few decades [10,11]. Therefore, in order to create a pleasant residential environment, it is necessary to develop policies to accurately identify images. After installing a Micasense RedEdge MX sensor on the UAV, a UAV flight plan was established using DJI GS Pro S/W, and multispectral images were acquired through UAV photography.
Next, by merging images using Pix4D Mapper S/W, orthomosaic images were constructed for the blue, green, red, and near-infrared bands. By combining the images for each band using ArcGIS S/W, various vegetation indices, such as the NDVI, GNDVI, BNDVI, RGBVI, GRVI, and SAVI, were calculated.
In addition, by comparing and analyzing the vegetation index values for various land covers in the target area, points showing abnormal values were identified, and sections to be considered when analyzing vegetation in urban areas were reviewed. Finally, a novel vegetation index was derived to effectively analyze the vegetation distribution area reflecting the urban land cover.  Figure 2 shows the part of the Vision College of Jeonju-si, located in Jeollabuk-do, Republic of Korea, that was selected as the target site of this study. The area contains vegetation areas, such as trees and lawns, as well as non-vegetation areas, such as an artificial First, the study site was selected, and eight ground control points (GCPs) were determined through a Virtual Reference Station (VRS) survey as reference points for the UAV images. After installing a Micasense RedEdge MX sensor on the UAV, a UAV flight plan was established using DJI GS Pro S/W, and multispectral images were acquired through UAV photography.

Study Area
Next, by merging images using Pix4D Mapper S/W, orthomosaic images were constructed for the blue, green, red, and near-infrared bands. By combining the images for each band using ArcGIS S/W, various vegetation indices, such as the NDVI, GNDVI, BNDVI, RGBVI, GRVI, and SAVI, were calculated.
In addition, by comparing and analyzing the vegetation index values for various land covers in the target area, points showing abnormal values were identified, and sections to be considered when analyzing vegetation in urban areas were reviewed. Finally, a novel vegetation index was derived to effectively analyze the vegetation distribution area reflecting the urban land cover. Figure 2 shows the part of the Vision College of Jeonju-si, located in Jeollabuk-do, Republic of Korea, that was selected as the target site of this study. The area contains vegetation areas, such as trees and lawns, as well as non-vegetation areas, such as an artificial grass playground, playground tracks, urethane basketball courts, long-run steel roofing, and waterproof coated roofs.  Figure 3a shows the DJI Matrice 210 UAV equipped with the Micasense RedEdge MX sensor, which was used to obtain multispectral images. The Micasense RedEdge MX sensor is widely used to analyze vegetation, agricultural management practices, and water quality because it can acquire blue, green, red, and near-infrared spectral images [51][52][53][54]. Its resolution is about 8 cm at an altitude of 120 m, providing very precise image information, and its weight is 231.9 g, which is very light [55], so it has the advantage of allowing sufficient flight time for the DJI Matrice 210, as shown in Table 1.

UAV Image Processing
The DJI GS Pro S/W was used to establish the UAV flight plan, as shown in Figure 4. The degree of longitudinal and lateral redundancy were designed to be 80% and 70%, respectively, and photography was performed at a flight altitude of about 100 m to acquire images with a resolution of 6 cm. Radiometric calibration was performed before the flight using a calibrated reflectance panel (CRP) and the incident light sensor of the Micasense RedEdge MX camera.  Figure 3a shows the DJI Matrice 210 UAV equipped with the Micasense RedEdge MX sensor, which was used to obtain multispectral images. The Micasense RedEdge MX sensor is widely used to analyze vegetation, agricultural management practices, and water quality because it can acquire blue, green, red, and near-infrared spectral images [51][52][53][54]. Its resolution is about 8 cm at an altitude of 120 m, providing very precise image information, and its weight is 231.9 g, which is very light [55], so it has the advantage of allowing sufficient flight time for the DJI Matrice 210, as shown in Table 1.    The DJI GS Pro S/W was used to establish the UAV flight plan, as shown in Figure 4. The degree of longitudinal and lateral redundancy were designed to be 80% and 70%, respectively, and photography was performed at a flight altitude of about 100 m to acquire images with a resolution of 6 cm. Radiometric calibration was performed before the flight using a calibrated reflectance panel (CRP) and the incident light sensor of the Micasense RedEdge MX camera.    To secure the positional accuracy of the UAV images, eight GCPs were selected, as shown in Table 2, and the transverse Mercator (TM) coordinates of the GRS80 ellipsoid were acquired through the VRS survey. Each captured image has Universal Transverse Mercator (UTM) coordinates on the WGS84 ellipsoid, obtained through a Global Positioning System (GPS) receiver attached to the UAV. To convert these coordinates into the TM coordinates of the GRS80 ellipsoid used in Korea, the survey results were matched to the GCPs. To secure the positional accuracy of the UAV images, eight GCPs were selected, as shown in Table 2, and the transverse Mercator (TM) coordinates of the GRS80 ellipsoid were acquired through the VRS survey. Each captured image has Universal Transverse Mercator (UTM) coordinates on the WGS84 ellipsoid, obtained through a Global Positioning System (GPS) receiver attached to the UAV. To convert these coordinates into the TM coordinates of the GRS80 ellipsoid used in Korea, the survey results were matched to the GCPs. In addition, images from each band were merged using Pix4D Mapper S/W, and index maps were constructed for each red, green, blue, and near-infrared image to calculate the vegetation index ( Figure 5). In addition, images from each band were merged using Pix4D Mapper S/W, and index maps were constructed for each red, green, blue, and near-infrared image to calculate the vegetation index ( Figure 5).

UAV Image Processing
The following are the general steps of Pix4D Mapper S/W image processing: (1) Image alignment and sparse point cloud building, (2)

Vegetation Analysis Using Existing Vegetation Indices
In this study, various existing vegetation indices were analyzed using ArcGIS S/W from the multispectral images constructed through UAV image processing (Table 3).  Table 4 gives the statistical characteristics of each vegetation index.

Vegetation Analysis Using Existing Vegetation Indices
In this study, various existing vegetation indices were analyzed using ArcGIS S/W from the multispectral images constructed through UAV image processing (Table 3). Figure 6 shows the vegetation index maps of NDVI, GNDVI, BNDVI, RGBVI, GRVI, and SAVI. Table 4 gives the statistical characteristics of each vegetation index.  The target area contains various types of land cover, such as trees, grass, an artificial grass playground, long-run steel roofing, waterproof coated roofs, and urethane-coated areas. Therefore, as illustrated in Figure 7, a representative location was selected for each type of land cover, and three sample points were chosen for each location. The target area contains various types of land cover, such as trees, grass, an artificial grass playground, long-run steel roofing, waterproof coated roofs, and urethane-coated areas. Therefore, as illustrated in Figure 7, a representative location was selected for each type of land cover, and three sample points were chosen for each location.
The UAV flight took place on 2 April 2020, in the early spring. At this time of year, the vegetation vitality of trees is high, but the vegetation vitality of grass is medium to low. Grass-covered positions showing medium vegetation vitality (Nos. [19][20][21] and slightly lower vitality (Nos. [22][23][24] were selected and compared.  Table 5 shows the results of the six existing vegetation indices for each type of land cover.
When vegetation was classified using NDVI, GNDVI, BNDVI, RGBVI, GRVI, and SAVI, there were locations in which the index values of long-run steel roofing, waterproof coated roofs, and urethane-coated areas were slightly higher than that of grass. Therefore, it was necessary to develop a novel vegetation index suitable for urban areas to effectively address this issue.  The UAV flight took place on 2 April 2020, in the early spring. At this time of year, the vegetation vitality of trees is high, but the vegetation vitality of grass is medium to low. Grass-covered positions showing medium vegetation vitality (Nos. [19][20][21] and slightly lower vitality (Nos. [22][23][24] were selected and compared. Table 5 shows the results of the six existing vegetation indices for each type of land cover.
When vegetation was classified using NDVI, GNDVI, BNDVI, RGBVI, GRVI, and SAVI, there were locations in which the index values of long-run steel roofing, waterproof coated roofs, and urethane-coated areas were slightly higher than that of grass. Therefore, it was necessary to develop a novel vegetation index suitable for urban areas to effectively address this issue.

A Novel Vegetation Index That Accurately Reflects Urban Land Cover
Thus, in urban areas, use of existing vegetation indices (NDVI, GNDVI, BNDVI, RGBVI, GRVI, and SAVI) results in misclassification of long-run steel roofing, urethanecoated, and waterproof coated areas as vegetation.
To address this problem, many equations were created by combining the blue, green, red, and nir bands, and three equations that can be used for urban vegetation analysis were finally selected through a manual process as Table 6. Finally, the optimal vegetation index was suggested by verifying the result of vegetation classification for each type of land cover.      To quantitatively verify the novel vegetation index proposed in this study, five validation points were selected for each type of land cover through field surveys. Among these points, twenty validation points in grass and trees were classified as vegetation, while thirty validation points in the areas with long-run steel roofing, artificial playgrounds, playground tracks, urethane coatings, and waterproof coated roofs were classified as non-vegetation. Figures 11-13 show the validation points for each type of land cover on the vegetation map created by applying the novel vegetation index equation. Table 8 shows the kappa coefficient for the comparison between the novel vegetation map and the ground truth results. The kappa coefficients of the squared Blue-Green NDVI index and squared Red-Green NDVI index were both 0.8, and the kappa coefficient of the squared Red-Blue NDVI index was the highest, at 1.0. Therefore, the squared Red-Blue NDVI index was found to be the most effective for vegetation analysis, reflecting the characteristics of urban land cover. First, using the squared Blue-Green NDVI index, the vegetation indices of the red urethane-coated areas (Nos. 13-15) ranged from 0.72 to 0.74, which were slightly higher than the values of lower-vitality grass (Nos. [22][23][24]. Therefore, as shown in Figure 8b, when 0.62 (the minimum value of lower-vitality grass) was set as the threshold value, the red urethane-coated areas were misclassified as vegetation.
Second, using the squared Red-Green NDVI index, the vegetation indices of the longrun steel roofing (Nos. 1-3) were between 0.73 and 0.78, which were similar to the values of the medium-vitality grass (Nos. [19][20][21] and higher than those of the lower-vitality grass (Nos. [22][23][24]. Therefore, as shown in Figure 9b, when 0.48 (the minimum value of the lower-vitality grass) was set as the threshold value, the areas with long-run steel roofing were misclassified as vegetation.
Finally, using the squared Red-Blue NDVI index, the vegetation indices of the long-run steel roofing (Nos. 1-3), urethane-coated (Nos. [10][11][12][13][14][15], and waterproof-coated (Nos. [16][17][18] areas were all lower than the vegetation index of lower-vitality grass (Nos. [22][23][24]. Therefore, as shown in Figure 10b, when 0.61 (the minimum value of the lower-vitality grass) was set as the threshold value, only areas with trees and grass were classified as vegetation areas, see Table 7. To quantitatively verify the novel vegetation index proposed in this study, five validation points were selected for each type of land cover through field surveys. Among these points, twenty validation points in grass and trees were classified as vegetation, while thirty validation points in the areas with long-run steel roofing, artificial playgrounds, playground tracks, urethane coatings, and waterproof coated roofs were classified as non-vegetation. Figures 11-13 show the validation points for each type of land cover on the vegetation map created by applying the novel vegetation index equation. Table 8 shows the kappa coefficient for the comparison between the novel vegetation map and the ground truth results. The kappa coefficients of the squared Blue-Green NDVI index and squared Red-Green NDVI index were both 0.8, and the kappa coefficient of the squared Red-Blue NDVI index was the highest, at 1.0. Therefore, the squared Red-Blue NDVI index was found to be the most effective for vegetation analysis, reflecting the characteristics of urban land cover. Appl. Sci. 2021, 11, x FOR PEER REVIEW 14 of 18       Existing studies using UAV have focused mainly on analyzing vegetation and topography in urban areas [41][42][43][44][45][46]. There have been studies to detect green-roof using the supervised classification method based on UAV images, but not to develop the vegetation index considering green-roof [41]. Compared with the existing UAV studies, the differences of the results performed in this study were added to the text. Therefore, the new index developed in this study will be a very effective method for classifying vegetation in urban areas with various land covers.

Conclusions
Accurate vegetation data are vital in various fields, such as urban planning, landscaping, and water quality analysis. In recent years, in order to effectively promote urban regeneration projects, research efforts have attempted to quantitatively analyze the distribution of vegetation in cities using satellite images or UAV images. In particular, vegetation indices, such as NDVI, are widely used for vegetation analysis using remote sensing technologies, such as satellite images and UAV images.  Existing studies using UAV have focused mainly on analyzing vegetation and topography in urban areas [41][42][43][44][45][46]. There have been studies to detect green-roof using the supervised classification method based on UAV images, but not to develop the vegetation index considering green-roof [41]. Compared with the existing UAV studies, the differences of the results performed in this study were added to the text. Therefore, the new index developed in this study will be a very effective method for classifying vegetation in urban areas with various land covers.

Conclusions
Accurate vegetation data are vital in various fields, such as urban planning, landscaping, and water quality analysis. In recent years, in order to effectively promote urban regeneration projects, research efforts have attempted to quantitatively analyze the distribution of vegetation in cities using satellite images or UAV images. In particular, vegetation indices, such as NDVI, are widely used for vegetation analysis using remote sensing technologies, such as satellite images and UAV images.
Various types of land cover, such as long-run steel roofing, waterproof coated roofs, and urethane-coated areas, exist in urban areas, and this study found that existing vegetation indices assign inappropriately high values to such areas. Therefore, in this study, existing vegetation indices, such as NDVI, GNDVI, BNDVI, RGBVI, GRVI, and SAVI, were used to quantitatively analyze the vegetation index values of different types of land cover. This was done based on high-resolution multispectral images taken by a UAV equipped with a Micasense RedEde MX camera.
As mentioned above, the vegetation index values of long-run steel roofing, waterproof coated roofs, and urethane-coated areas were similar to or higher than those of grass with a slightly low degree of vitality. To address the problem of misclassification according to the characteristics of urban land cover, a test was conducted by combining multiple spectral bands. Finally, three equations were selected and analyzed.
According to a kappa coefficient analysis, the squared Red-Blue NDVI index, which was created by combining information from the blue, red, and near-infrared bands, most accurately reflected the characteristics of urban land cover. The novel vegetation index presented in this study is expected to be very useful in analyzing the vegetation in regions with various types of land cover when conducting urban regeneration projects in the future.