Detection of Aquatic Plants Using Multispectral UAV Imagery and Vegetation Index

: In this study, aquatic plants in a small reservoir were detected using multispectral UAV (Unmanned Aerial Vehicle) imagery and various vegetation indices. A Fireﬂy UAV, which has both ﬁxed-wing and rotary-wing ﬂight modes, was ﬂown over the study site four times. A RedEdge camera was mounted on the UAV to acquire multispectral images. These images were used to analyze the NDVI (Normalized Di ﬀ erence Vegetation Index), ENDVI (Enhance Normalized Di ﬀ erence Vegetation Index), NDREI (Normalized Di ﬀ erence RedEdge Index), NGRDI (Normalized Green-Red Di ﬀ erence Index), and GNDVI (Green Normalized Di ﬀ erence Vegetation Index). As for multispectral characteristics, waterside plants showed the highest reﬂectance in R nir , while ﬂoating plants had a higher reﬂectance in R re . During the hottest season (on 25 June), the vegetation indices were the highest, and the habitat expanded near the edge of the reservoir. Among the vegetation indices, NDVI was the highest and NGRDI was the lowest. In particular, NGRDI had a higher value on the water surface and was not useful for detecting aquatic plants. NDVI and GNDVI, which showed the clearest di ﬀ erence between aquatic plants and water surface, were determined to be the most e ﬀ ective vegetation indices for detecting aquatic plants. Accordingly, the vegetation indices using multispectral UAV imagery turned out to be e ﬀ ective for detecting aquatic plants. A further study will be accompanied by a ﬁeld survey in order to acquire and analyze more accurate imagery information.


Introduction
It is widely known that aquatic plants in water bodies such as rivers and reservoirs can improve water quality [1][2][3]. In particular, floating plants live on the water surface rather than rooting in soil; as their leaves and stems float on the water surface, floating plants block light penetration, hinder the photosynthesis of phytoplankton, and reduce the concentration of chlorophyll-a, thereby suppressing green algae [4]. Recently, the ability of aquatic plants to improve water quality has been utilized to purify polluted rivers or reservoirs from an ecological perspective [5], and various studies are being conducted to investigate the distribution of aquatic plants living in water.
Field surveys have been the conventional method for identifying vegetation. The salient advantage of this method is the careful observation of plant species. However, if a large area is investigated, then field surveys require a large amount of labor, cost, and time [6,7]. In particular, as aquatic plants often live in inaccessible regions, direct field surveys are not as effective as expected.
surveys. As satellites can periodically provide imagery while moving in an orbit, they are very useful for analyzing patterns of ecosystem change [8,9]. Furthermore, a wide area stretching from a few meters to kilometers can be analyzed within a short time [10]. Nevertheless, as most satellite images have a middle or low spatial resolution, they are not sufficient for the precise analysis of plant species or small-scale ecosystems [11][12][13]. To solve this problem, UAVs are actively used for surveys [14,15]. Since UAVs are operated at lower altitudes than satellites, higher resolution images can be acquired [16,17], enabling plant species or habitats to be precisely analyzed [18,19]. Moreover, UAVs can be operated to freely acquire information about any area at any time desired, and can be effectively used to monitor inaccessible regions, for which satellite imagery cannot be acquired; additionally, they can identify the impact of environmental factors like heavy rainfall, snowfall, and heat waves [20].
Vegetation surveys using UAVs mainly obtain high resolution images and analyze the unique spectral information of plants to classify plant species or identify the distribution of habitats [7,13,16,[21][22][23][24]. For such an analysis, a vegetation index (VI) is usually applied, which is derived from the characteristics of various spectral wavelengths [25][26][27]. The vegetation index is an indicator that quantifies the presence or absence of vegetation, the greenness of land cover, or the vitality of vegetation. Currently, the vegetation index is utilized not only for monitoring plant habitats, but also in various areas including the analysis of drought and environmental change [28,29]. However, most previous studies have applied the vegetation index to inland forest or agricultural areas, whereas only a few studies have dealt with aquatic plants. This is important because, as aquatic plants float on the water surface, they may have different spectral properties than those of land plants.
This study evaluated a UAV-based method of surveying the distribution of aquatic plant habitats in a small reservoir located in Gyeongju-si, Republic of Korea. The analysis proceeded as follows: (1) multispectral UAV images were acquired and the spectral characteristics for aquatic plants and water surface were analyzed, and (2) various vegetation indices used in the existing studies were analyzed to see what method was appropriate for surveying the habitats of aquatic plants.

Study Reservoir
This study was conducted for the region of Hayeombulji (N35.801, E129.186), located in Gyeongju-si, Gyeongsangbuk-do, Republic of Korea ( Figure 1). Hayeombulji has an area of 47.103 m 2 , a diameter of about 230 m, and a depth of 4~5 m [30]. With regard to the climate, the reservoir has four distinct seasons. In the summer, during which a large amount of agricultural activity takes place, the temperature reaches up to approximately 30 • C, and the reservoir water is extensively used for agriculture. Accordingly, aquatic plant habitats need to be analyzed in order to ensure a supply of clean water for agricultural purposes.

UAV Imagery Acquisition and Processing
In this study, a RedEdge multispectral camera from MicaSense was mounted on a Firefly to obtain UAV imagery ( Figure 2). The Firefly uses both rotary-wing and fixed-wing flight modes. The rotary-wing mode is adopted during the taking-off and landing phases, which enables the vertical movement of the UAV. After take-off, the UAV flies horizontally in the fixed-wing flight mode, during which imagery is acquired. Consequently, the Firefly has relatively long flight times, thereby addressing a disadvantage of rotary-wing UAVs as well as requiring relatively little space for takeoff and landing, overcoming a disadvantage of fixed-wing UAVs. The MicaSense RedEdge multispectral camera (MicaSense Inc., Seattle, WA, USA) has five spectral wavelength bands (blue, green, red, near IR, RedEdge). The detailed specifications of the Firefly UAV and the MicaSense RedEdge multispectral camera are presented in Tables 1-3.

UAV Imagery Acquisition and Processing
In this study, a RedEdge multispectral camera from MicaSense was mounted on a Firefly to obtain UAV imagery ( Figure 2). The Firefly uses both rotary-wing and fixed-wing flight modes. The rotary-wing mode is adopted during the taking-off and landing phases, which enables the vertical movement of the UAV. After take-off, the UAV flies horizontally in the fixed-wing flight mode, during which imagery is acquired. Consequently, the Firefly has relatively long flight times, thereby addressing a disadvantage of rotary-wing UAVs as well as requiring relatively little space for take-off and landing, overcoming a disadvantage of fixed-wing UAVs. The MicaSense RedEdge multispectral camera (MicaSense Inc., Seattle, WA, USA) has five spectral wavelength bands (blue, green, red, near IR, RedEdge). The detailed specifications of the Firefly UAV and the MicaSense RedEdge multispectral camera are presented in Tables 1-3.

UAV Imagery Acquisition and Processing
In this study, a RedEdge multispectral camera from MicaSense was mounted on a Firefly to obtain UAV imagery ( Figure 2). The Firefly uses both rotary-wing and fixed-wing flight modes. The rotary-wing mode is adopted during the taking-off and landing phases, which enables the vertical movement of the UAV. After take-off, the UAV flies horizontally in the fixed-wing flight mode, during which imagery is acquired. Consequently, the Firefly has relatively long flight times, thereby addressing a disadvantage of rotary-wing UAVs as well as requiring relatively little space for takeoff and landing, overcoming a disadvantage of fixed-wing UAVs. The MicaSense RedEdge multispectral camera (MicaSense Inc., Seattle, WA, USA) has five spectral wavelength bands (blue, green, red, near IR, RedEdge). The detailed specifications of the Firefly UAV and the MicaSense RedEdge multispectral camera are presented in Tables 1-3.     UAV imagery was acquired four times (25 April, 20 May, 25 June, and 24 October 2018). The flight path was set using Firefly 6 Planner, which is the dedicated software for the Firefly UAV, and considered geographic features around the study area ( Figure 3). During the first flight on 25 April 2018, the flight path was set, and the UAV was operated in the same path for the remaining flights. While the UAV was operated, the duplication rate, flight time, and flight altitude were set to 80%, about 20 min, and 150 m, respectively. Before each flight, the reflectance panel was used to calibrate the reflectance of the multispectral imagery. The operation of the UAV was initiated by pressing the start button in Firefly 6 Planner. With the UAV flight, 338 images were acquired for each spectral wavelength band. Five orthoimages were fabricated in each spectral wavelength band from images obtained from the UAV by using Pix4D Software 2.1 (Pix4D: Lausanne, Switzerland, 2016). The analysis process consists of orthoimages through initial processing, point cloud, mesh generation, orthomosaic, and image correction using reflection panel information.

Vegetation Index
As shown in Table 2, a total of five vegetation indices were analyzed. These indices were NDVI (normalized difference vegetation index), ENDVI (extended normalized difference vegetation index), NDREI (normalized difference red edge index), NGRDI (normalized green-red difference index), and GNDVI (green normalized difference vegetation index). Five vegetation indices were selected because they are widely used for extracting vegetation areas in previous studies and can be analyzed using the spectral wavelength bands acquired from the RedEdge M sensor. For the analysis method, the Raster Calculator Tool in ArcGIS Ver. 10.2 was applied to the orthoimages of each spectral wavelength band, as shown in Table 4.
Remote Sens. 2020, 12, 387 5 of 16 the reflectance of the multispectral imagery. The operation of the UAV was initiated by pressing the start button in Firefly 6 Planner. With the UAV flight, 338 images were acquired for each spectral wavelength band. Five orthoimages were fabricated in each spectral wavelength band from images obtained from the UAV by using Pix4D Software 2.1 (Pix4D: Lausanne, Switzerland, 2016). The analysis process consists of orthoimages through initial processing, point cloud, mesh generation, orthomosaic, and image correction using reflection panel information.

Vegetation Index
As shown in Table 2, a total of five vegetation indices were analyzed. These indices were NDVI (normalized difference vegetation index), ENDVI (extended normalized difference vegetation index), NDREI (normalized difference red edge index), NGRDI (normalized green-red difference index), and GNDVI (green normalized difference vegetation index). Five vegetation indices were selected because they are widely used for extracting vegetation areas in previous studies and can be analyzed  Table 4. Vegetation indices using multi-spectral imagery.

Name
Abbreviation Formula Value Range Reference

Profile Analysis
A profile analysis was performed to identify the characteristics of spectral wavelength bands and vegetation indices with respect to the covering characteristics of water space such as aquatic plants and water surface. Polyline features were created in ArcGIS 10.2 at 1 m intervals for some sections of the study area, which included both aquatic plants and the water surface ( Figure 4), while the reflectance of a spectral wavelength band and the mean vegetation index for each section were calculated using zonal statistics tools. Subsequently, the mean values were calculated for the polyline features, and the reflectance of the spectral wavelength band and the characteristics of vegetation indices were analyzed according to the presence of aquatic plants.
plants and water surface. Polyline features were created in ArcGIS 10.2 at 1 m intervals for some sections of the study area, which included both aquatic plants and the water surface ( Figure 4), while the reflectance of a spectral wavelength band and the mean vegetation index for each section were calculated using zonal statistics tools. Subsequently, the mean values were calculated for the polyline features, and the reflectance of the spectral wavelength band and the characteristics of vegetation indices were analyzed according to the presence of aquatic plants.  Figure 5 illustrates the reflectance of each spectral wavelength band, which was obtained through the profile analysis. On 25 April, the closer to location A (i.e., the reservoir edge), higher reflectance was shown in every wavelength band. In particular, Rnir and Rre had a higher reflectance than the remaining wavelength bands. In the other areas, there was no significant difference in terms of reflectance among the spectral wavelength bands. In contrast to the reflectance recorded on 25  Figure 5 illustrates the reflectance of each spectral wavelength band, which was obtained through the profile analysis. On 25 April, the closer to location A (i.e., the reservoir edge), higher reflectance was shown in every wavelength band. In particular, R nir and R re had a higher reflectance than the remaining wavelength bands. In the other areas, there was no significant difference in terms of reflectance among the spectral wavelength bands. In contrast to the reflectance recorded on 25 April, those recorded on both 20 May and 25 June displayed peaks in many areas. This was because the habitats of waterside plants and floating plants had expanded, which resulted in an increase of reflectance in areas including such habitats. In particular, the waterside plants had the highest reflectance in R nir , while the floating plants showed a high reflectance in R re . Accordingly, the waterside plants and the floating plants had different reflectance patterns in each spectral wavelength band (Figure 5a). Unlike the other spectral band, R r and R b did not show any significant change of reflectance. The waterside plants had a similar pattern in every wavelength band. On the other hand, the floating plants showed an increasing and decreasing pattern only in R nir , R re , and R g , but displayed no significant change in R r and R b . As the floating plants are submerged in water or afloat on water, they tend to be much more affected by spectral reflectance and thus may show a different characteristic of reflectance in each spectral wavelength band from the waterside plants that are not submerged in water. Consequently, the reflectance characteristics in each spectral wavelength band were clearly different according to the type of aquatic plants and habitats. Moreover, among the aquatic plants, the floating plants and the waterside plants showed a significant difference in terms of reflectance in each spectral wavelength band according to habitat. they tend to be much more affected by spectral reflectance and thus may show a different characteristic of reflectance in each spectral wavelength band from the waterside plants that are not submerged in water. Consequently, the reflectance characteristics in each spectral wavelength band were clearly different according to the type of aquatic plants and habitats. Moreover, among the aquatic plants, the floating plants and the waterside plants showed a significant difference in terms of reflectance in each spectral wavelength band according to habitat.   approximately 30 °C and plants proliferate; the aquatic plants of the reservoir followed this trend, and vegetation indices seemed to be highest in this season. The areas with a vegetation index of 0.4 or above tended to gradually expand from the edge to the interior of the reservoir, indicating that the habitats of aquatic plants spread in the same direction. In comparison to that in June, the vegetation indices recorded on October 24 drastically decreased; this corresponded with the gradual temperature decrease in fall (NDVI: 0.388, ENDVI: 0.361, NDREI: 0.148, NGRDI: 0.200, GNDVI: 0.203).  displayed a significant difference. The NDVI was calculated by using R nir and R r . As is clear from Figure 5, the aquatic plants had the highest reflectance in R nir and the lowest one in R r . In other words, the difference in terms of reflectance was largest between these two wavelength bands. Accordingly, the NDVI was higher than any other vegetation indices. Remote Sens. 2020, 12, x FOR PEER REVIEW 9 of 16 displayed a significant difference. The NDVI was calculated by using Rnir and Rr. As is clear from Figure 5, the aquatic plants had the highest reflectance in Rnir and the lowest one in Rr. In other words, the difference in terms of reflectance was largest between these two wavelength bands. Accordingly, the NDVI was higher than any other vegetation indices.

Results of Vegetation Index Analysis
On the other hand, the GNDVI and NDREI were low. On April 25 and on May 20, when the aquatic plants had not yet proliferated, GNDVI was the lowest (0.017 on April 25, 0.071 on May 20). On June 25 and on October 24, when the aquatic plants had proliferated, NDREI was the lowest (0.164 on June 25, 0.148 on October 24). The GNDVI was calculated based on Rnir and Rg. As the difference between Rnir and Rg was larger than the one between Rnir and Rre, which are used to calculate NDREI, the GNDVI was greater than the NDREI. As shown in Figure 6, most parts of the reservoir had a GNDVI of −0.2 or below on April 25 and on May 20. On the other hand, NDREI was between −0.2 and 0 over a large area. Accordingly, the mean value of the GNDVI was lower than that of NDREI. On June 25 and on October 24, many parts of the reservoir had a GNDVI of 0.4 or above, while only the reservoir edge had an increased NDREI, although the other areas did not.
Consequently, the vegetation indices had different trends according to the proliferation time of aquatic plants. This seems to be because of the characteristics of spectral wavelength bands, which were used to calculate the vegetation indices. Thus, such characteristics of spectral wavelength bands and vegetation indices need to be carefully analyzed and utilized in order to detect aquatic plants. Figure 8 presents the profile analysis results for each season and vegetation index. Generally, the vegetation indices had different trends according to the season. In areas where aquatic plants were detected, the NDVI was the highest, followed by the ENDVI and GNDVI. On April 25 and on May 20, the GNDVI was higher than the ENDVI. In contrast, on June 25 and on October 24, the ENDVI was higher than the GNDVI. The NGRDI of aquatic plants was low, but nevertheless higher than the NDVI, ENDVI, and GNDVI on the water surface. As shown in Figure 5, the reflectance did not significantly vary among the wavelength bands on the water surface. However, since the aquatic The GNDVI was calculated based on R nir and R g . As the difference between R nir and R g was larger than the one between R nir and R re , which are used to calculate NDREI, the GNDVI was greater than the NDREI. As shown in Figure 6, most parts of the reservoir had a GNDVI of −0.2 or below on April 25 and on May 20. On the other hand, NDREI was between −0.2 and 0 over a large area. Accordingly, the mean value of the GNDVI was lower than that of NDREI. On June 25 and on October 24, many parts of the reservoir had a GNDVI of 0.4 or above, while only the reservoir edge had an increased NDREI, although the other areas did not.

Characteristics of Aquatic Plant Vegetation Indices
Consequently, the vegetation indices had different trends according to the proliferation time of aquatic plants. This seems to be because of the characteristics of spectral wavelength bands, which were used to calculate the vegetation indices. Thus, such characteristics of spectral wavelength bands and vegetation indices need to be carefully analyzed and utilized in order to detect aquatic plants. Figure 8 presents the profile analysis results for each season and vegetation index. Generally, the vegetation indices had different trends according to the season. In areas where aquatic plants were detected, the NDVI was the highest, followed by the ENDVI and GNDVI. On April 25 and on May 20, the GNDVI was higher than the ENDVI. In contrast, on June 25 and on October 24, the ENDVI was higher than the GNDVI. The NGRDI of aquatic plants was low, but nevertheless higher than the NDVI, ENDVI, and GNDVI on the water surface. As shown in Figure 5, the reflectance did not significantly vary among the wavelength bands on the water surface. However, since the aquatic plants had a high reflectance in R nir and R re , the NGRDI, which is calculated using R g and R r , was low in the aquatic plants, but high on the water surface. plants had a high reflectance in Rnir and Rre, the NGRDI, which is calculated using Rg and Rr, was low in the aquatic plants, but high on the water surface.  A statistical analysis was performed to derive a vegetation index that could be used to detect aquatic plants, as shown in Table 5 Accordingly, NDVI and GNDVI seemed to be more effective than the other indices for detecting aquatic plants. When the NDVI and GNDVI were compared with each other in detail, there was no significant difference between the NDVI and GNDVI in different habitats on April 25. However, the GNDVI was lower than the NDVI by about 0.2 on the water surface. On May 20, the NDVI was higher than the GNDVI by about 0.05, but both showed a similar pattern across all habitats. However, the GNDVI was lower than the NDVI by about 0.1 on the water surface. On June 25, when the aquatic plants had proliferated, the NDVI was higher than the GNDVI by about 0.2, although both indices generally displayed a similar pattern. On 25 October, there was little difference between the vegetation indices between habitats. However, on the water surface, the GNDVI was lower than the NDVI by about 0.3; in fact, the NDVI was generally higher than the GNDVI. However, as for the difference between aquatic plants and the water surface, the GNDVI was higher than the NDVI. Figure 9 shows the analysis results of scatter plots with the criterion of the NDVI in order to identify vegetation index trends according to the characteristics of aquatic plants and the water surface. On 25 April, all the vegetation indices were positively (+) correlated to the NDVI, with the exception of the NGRDI. In particular, the ENDVI was very highly correlated to NDVI (R 2 = 0.9189). The NGRDI was negatively (−) correlated to the NDVI, but the correlation was very low (R 2 = 0.1306). On 20 May, all the vegetation indices were positively correlated to the NDVI. However, the R 2 of the NGRDI was as low as 0.1181. On 25 June, the R 2 was at least 0.4 in every vegetation index. In particular, the GNDVI and ENDVI showed a high correlation (GNDVI = 0.9168, ENDIV = 0.9715). Like the results for 25 April, all the vegetation indices were positively (+) correlated to the NDVI on 25 October, with the exception of NGRDI.

Characteristics of Aquatic Plant Vegetation Indices
The GNDVI, ENDVI, and NDREI were highly correlated to the NDVI and showed the same characteristics according to the covering types of the reservoir including aquatic plants and the water surface. On the other hand, the NGRDI had a low correlation to the NDVI and the negative correlation, that is, the opposite tendency was shown in all seasons except for summer (25 June), where there was a large amount of vegetation. This was because the NDVI was high in habitats, but the NGRDI was low in aquatic plants and rather high on the water surface.

Discussion
As a result of analyzing the characteristics of aquatic plants for various vegetation indices extracted by multispectral UAV imagery, all vegetation indices except the NGRDI showed a similar tendency, according to the presence or absence of aquatic plants. Among the existing studies using the NGRDI, [36] calculated the NGRDI by utilizing the UAV imagery information and analyzed the correlation between the biomass on the ground surface and the leaf area index (LAI). They reported that the correlation coefficient (r) between the NGRDI and biomass was 0.54 ~ 0.78; furthermore, the NGRDI from UAV imagery was cost-effective in predicting biomass. Additionally, [37] reported that the NGRDI was more effective than NDVI in extracting N fertilization. Furthermore, [38] revealed that NGRDI was effective in detecting green algae and biomass. Rasmussen et al. [25], Smigaj et al. [39], Shimada et al. [40], and Li et al. [41] also utilized the NGRDI to detect fertilizer or biomass in agricultural regions. As shown in the existing studies, since the NGRDI is derived by using the spectral wavelength bands of Rr and Rg, it seems to be effective in detecting green algae or analyzing the characteristics of plants in summer, where leaves are green due to the high vitality of plants. However, Rr and Rg did not have a significant difference with regard to the reflectance between

Discussion
As a result of analyzing the characteristics of aquatic plants for various vegetation indices extracted by multispectral UAV imagery, all vegetation indices except the NGRDI showed a similar tendency, according to the presence or absence of aquatic plants. Among the existing studies using the NGRDI, [36] calculated the NGRDI by utilizing the UAV imagery information and analyzed the correlation between the biomass on the ground surface and the leaf area index (LAI). They reported that the correlation coefficient (r) between the NGRDI and biomass was 0.54~0.78; furthermore, the NGRDI from UAV imagery was cost-effective in predicting biomass. Additionally, [37] reported that the NGRDI was more effective than NDVI in extracting N fertilization. Furthermore, [38] revealed that NGRDI was effective in detecting green algae and biomass. Rasmussen et al. [25], Smigaj et al. [39], Shimada et al. [40], and Li et al. [41] also utilized the NGRDI to detect fertilizer or biomass in agricultural regions. As shown in the existing studies, since the NGRDI is derived by using the spectral wavelength bands of R r and R g , it seems to be effective in detecting green algae or analyzing the characteristics of plants in summer, where leaves are green due to the high vitality of plants.
However, R r and R g did not have a significant difference with regard to the reflectance between aquatic plants and the water surface. Accordingly, the NGRDI does not seem to be appropriate for detecting aquatic plants.
Among the vegetation indices, except the NGRDI, the most effective vegetation indices of the aquatic plant areas were the NDVI and GNDVI. The NDVI and GNDVI were clearly distinguished as their values were most significantly different between aquatic plants and the water surface ( Figure 8). As shown in Figure 10, in order to see whether the NDVI and GNDVI can be used to detect aquatic plants, the NDVI and GNDVI were set to a specific value or above (NDVI: 0.35 or above, GNDVI: 0.25 or above) on 25 June when aquatic plants had proliferated, and the habitats of aquatic plants were extracted. In comparison with the RGB images, the distribution of habitats thus extracted was similar to the real distribution of aquatic plants, although there was a difference with regard to the aquatic plant area (i.e., area B). Additionally, the area of aquatic plant by the NDVI was found to be larger than that of the GNDVI (NDVI: 356.63 m 2 , GNDVI: 269.06 m 2 ).
Remote Sens. 2020, 12, x FOR PEER REVIEW 13 of 16 aquatic plants and the water surface. Accordingly, the NGRDI does not seem to be appropriate for detecting aquatic plants.
Among the vegetation indices, except the NGRDI, the most effective vegetation indices of the aquatic plant areas were the NDVI and GNDVI. The NDVI and GNDVI were clearly distinguished as their values were most significantly different between aquatic plants and the water surface ( Figure  8). As shown in Figure 10, in order to see whether the NDVI and GNDVI can be used to detect aquatic plants, the NDVI and GNDVI were set to a specific value or above (NDVI: 0.35 or above, GNDVI: 0.25 or above) on 25 June when aquatic plants had proliferated, and the habitats of aquatic plants were extracted. In comparison with the RGB images, the distribution of habitats thus extracted was similar to the real distribution of aquatic plants, although there was a difference with regard to the aquatic plant area (i.e., area B). Additionally, the area of aquatic plant by the NDVI was found to be larger than that of the GNDVI (NDVI: 356.63 m 2 , GNDVI: 269.06 m 2 ). These results show that it difficult to identify which vegetation indices are more effective in detecting aquatic plant areas. This is because the aquatic plant area in the reservoir is unknown These results show that it difficult to identify which vegetation indices are more effective in detecting aquatic plant areas. This is because the aquatic plant area in the reservoir is unknown accurately. However, the vegetation indices extracted from the UAV, which has been used for forest ecosystems, can be used to monitor the habitat change of aquatic ecosystems in high resolution. In addition, more accurate vegetation index values need to be applied in order to extract aquatic plant areas. Additionally, the vegetation index value, according to the species of aquatic plants, should be identified, which will need to be studied continuously in the future.

Conclusions
This study analyzed a method of detecting aquatic plants by utilizing multispectral UAV imagery and various vegetation indices for a small reservoir. The results of this study can be summarized as follows.
The vegetation indices were highest on 25 June, which corresponded to the hottest season, and lowest on 25 April. The seasonal change of vegetation indices revealed that the habitats of aquatic plants gradually expanded from the edge of the reservoir. When the vegetation indices were compared, the NDVI had the highest mean value of 0.820 on 25 June, and the difference in relation to the remaining vegetation indices was as large as approximately 0.2~0.4. On the other days, the NDVI was also the highest, but the difference between the NDVI and the remaining indices was only about 0.1. This is attributable to the reflectance characteristics of each spectral wavelength band. In other words, the NDVI is calculated by using R nir and R r , which have the largest difference in regard to reflectance. On the other hand, the GNDVI and NDREI belonged to the lowest group. The GNDVI was lowest on 25 April and on 20 May, when aquatic plants had not yet proliferated. The NDREI was lowest on 25 June, when aquatic plants had proliferated.
The profile analysis showed that the NDVI and GNDVI were effective for detecting aquatic plants since these indices had a larger deviation than the other indices. In addition, the NDVI and GNDVI had a similar tendency according to the covering characteristics of the reservoir. However, GNDVI was much lower than the NDVI on the water surface, indicating a clearer difference between the aquatic plants and the water surface. Except on 25 June, the GNDVI was similar to the NDVI in aquatic plants, but was much lower than the NDVI on the water surface.
Accordingly, for the UAV-based detection of aquatic plants, the GNDVI seems to be effective along with the NDVI, which has been widely applied in previous studies. However, when the habitats of aquatic plants are extracted using the NDVI and GNDVI, accurate vegetation index values of aquatic plants need to be identified. Although this study analyzed the characteristics of spectral information regarding aquatic plants using multispectral imagery, a field survey for verification and calibration is also necessary for the more accurate identification of spectral characteristics.
This study has demonstrated that UAV imagery is sufficiently applicable for analyzing the distribution of aquatic plants. This method is costly, so the necessity of a cost-effective method for utilizing UAVs has received much attention. With this regard, the above results indicate a low cost and high effective method for detecting aquatic plants. The method proposed in this study seems to be applicable to the area of surveying aquatic plants. Of course, it is still necessary to verify the accuracy of both imagery information and analysis results through a field survey. A further study will deal with this issue and supplement the limitations of this study.
Author Contributions: B.S. and K.P. conceived and designed the research; B.S. performed the field measurements; and K.P. and B.S. analyzed the spatial data and wrote the paper. All authors have read and agreed to the published version of the manuscript.