ANALYSIS OF FIELD SEAGRASS PERCENT COVER AND ABOVEGROUND CARBON STOCK DATA FOR NON-DESTRUCTIVE ABOVEGROUND SEAGRASS CARBON STOCK MAPPING USING WORLDVIEW-2 IMAGE

Remote sensing can make seagrass aboveground carbon stock (AGCseagrass) information spatially extensive and widely available. Therefore, it is necessary to develop a rapid approach to estimate AGCseagrass in the field to train and assess its remote sensing-based mapping. The aim of this research is to (1) analyze the Percent Cover (PCv)-AGCseagrass relationship in seagrass at the species and community levels to estimate AGCseagrass from PCv and (2) perform AGCseagrass mapping at both levels using WorldView-2 image and assess the accuracy of the resulting map. This research was conducted in Karimunjawa and Kemujan Islands, Indonesia. Support Vector Machine (SVM) classification was used to map seagrass species composition, and stepwise regression was used to model AGCseagrass using deglint, water column corrected, and principle component bands. The results were a rapid AGCseagrass estimation using an easily measured parameter, the seagrass PCv. At the community level, the AGCseagrass map had 58.79% accuracy (SEE = 5.41 g C m), whereas at the species level, the accuracy increased for the class Ea (64.73%, SEE = 6.86 g C m) and EaThCr (70.02%, SEE = 4.32 g C m) but decreased for ThCr (55.08%, SEE = 2.55 g C m). The results indicate that WorldView-2 image reflectance can accurately map AGCseagrass in the study area in the range of 15‒20 g C m for Ea, 10‒15 g C m for EaThCr, and 4‒8 g C m for ThCr. Based on our model, the AGCseagrass in the study area was estimated at 13.39 t C.


INTRODUCTION
The integration of remote sensing and field data for mapping aboveground carbon stock of seagrass (AGCseagrass) requires corresponding field data to train the regression model and assess the accuracy of the resulting map (Hossain et al., 2015;Tamondong et al., 2018). However, the process of obtaining this data is destructive, costly, and time-consuming (Misbari, 2017); hence, it is necessary to develop a fast and non-destructive approach to estimate AGCseagrass for use in training and validating remote sensing-based AGC mapping.
Equations to approximate AGC from easily measured parameters, such as percent cover, are currently available (Wahyudi et al., 2020) and are especially useful for AGCseagrass appraisal without considering species uniqueness. However, this study proposed a new alternative for predicting AGCseagrass using percent cover (PCv) while also factoring in the unique PCv-AGCseagrass relationship of each species. In addition to developing a species-specific PCv-AGCseagrass equation, we also integrate WorldView-2 multispectral remote sensing images with in situ AGCseagrass data, calculated using the PCv-AGCseagrass equation, to map AGCseagrass.
To achieve these objectives, this study measured seagrass PCv in the field and AGCseagrass in the laboratory and analyzed their relationship. We acknowledge that the interaction between 1 seagrass beds with a variety of seagrass species and distribution patterns. Since 1999, they have been part of the Karimunjawa National Park under the authority of the Indonesian Ministry of Forestry. Several seagrass species found in the study area are Enhalus acoroides (Ea), Halophila ovalis (Ho), Thalassia hemprichii (Th), Cymodocea rotundata (Cr), Cymodocea serrulata (Cs), Halodule pinifolia (Hp), Halodule uninervis (Hu), Syringodium isoetifolium (Si), and Thalassodendron ciliatum (Tc). In all parts of the study area, seagrass inhabits a narrow fringe along the shoreline. However, the western sides of the island have more seagrass than the eastern side because they have relatively weak waves, are surrounded or sheltered by smaller islands, have reef crests that decrease the incoming wave energies, and comprise a broader area of reef flats and shallow lagoons. These conditions are more favorable for seagrass growth.

Image Data
The remote sensing image used in this research project was WorldView-2 (WV2) LV3X that had been orthorectified and thus did not require geometric correction. This project used the visible bands of WV2 images for mapping AGCseagrass and the nearinfrared bands to perform the sunglint correction. There are time differences between the WV2 image acquisition (24 May 2012) and the field data collection (July 2011, April 2012, and November 2012. Because of seasonal seagrass patterns in the study area, there would be changes in seagrass abundance and biophysical properties composition during these times, which may influence the empirical model quality and the AGCseagrass mapping accuracy. However, changes due to human disturbances are believed to be low since most seagrass areas are within the management of Karimunjawa National Park.

Seagrass sample analysis:
In each sampling site, seagrass leaves were harvested in a quadrat of 1 × 1 m (1 m 2 ) to standardize the measurement unit of biophysical properties. This process also took pictures of the quadrat overlaid on seagrass and recorded the coordinate of each sample. In total, 45 seagrass samples were collected. Then, the PCv and AGCseagrass were measured for each seagrass species present in the quadrat, i.e., Ea, Th, Cr, Hu, Si, and Ho. The PCv, which refers to the projected percentage of horizontal seagrass coverage within the quadrat, was estimated based on the seagrass species identified in the quadrat photos. In this research, the PCv was measured when the seagrass is submerged. The carbon stock of each species was measured in the laboratory following the proximate analysis procedure, and the AGCseagrass in each quadrant was also calculated for each species.
These samples were used to analyze the relationship between PCv and AGCseagrass at community and species levels. Regression analyses were performed first at the quadrat level consisting of single or mixed seagrass species to determine the relationship between PCv and AGCseagrass at the community level and then for each species to understand their unique PCv-AGCseagrass relationship.

Seagrass photo-quadrat analysis:
To obtain a representative number of samples to train the regression model and validate the resulting map, the distribution of seagrass samples must represent variations in seagrass conditions in the study area. This research project used a stratified random, aligned sampling to determine the photo-quadrat sample locations. The mapping unit used to determine the sampling locations in a stratified random manner was seagrass density, whose variations were identified visually on true-color-composite (TCC) WV2 image (R-G-B 5-3-2) using two interpretation keys: tone and texture. Seagrass pixels with darker tones indicate higher density and vice versa. A cluster of pixels of higher seagrass density has a smoother texture than those of lower density, where the background in between seagrasses increases the reflectance variations, roughening the texture. The location of each sample was center-aligned in the sampling's mapping unit to minimize the risk of collecting samples from adjacent units due to the combined errors of GPS and image's geometric accuracy. Afterward, the PCv of each seagrass species in each photoquadrat sample was interpreted and then converted to AGCseagrass based on the resultant regression function showing the PCv-AGCseagrass relationship in seagrass.
The training areas and validation samples were carefully selected by considering the spatial distributions of sample locations, species dynamics, and seagrass PCv differences so that both training area and validation samples cover similar seagrass variations. Randomly selecting the samples may lead to unequal distribution of the training areas and validation samples. Figure  1 shows the location of the study area and the field data distribution.

Image Corrections
The procedure to convert the Digital Number (DN) of WV2 images to TOA radiance (LTOA) and TOA reflectance (RTOA) and the parameters involved, i.e., absolute radiometric calibration factors, mean solar exoatmospheric irradiance (ESUN), and zenith angle, were described in (Updike and Comp, 2010). The atmospheric path radiance was removed from Top-of-Atmosphere (TOA) reflectance image using the Dark-Object Subtraction (DOS) algorithm (Chavez et al., 1977) to obtain Bottom-of-Atmosphere (BOA) reflectance (RBOA) image. The DOS method was selected because it is simple but with performance comparable to more robust methods, such as Fast Line-of-Sight Atmospheric Analysis of Hypercubes (FLAASH) and 6S radiative transfer models for multispectral images (Lantzanakis et al., 2016;Wicaksono, Hafizt, 2018).
Sunglint-free reflectance (RD) was obtained using a method developed by (Hedley et al., 2005). Principle Component Analysis (PCA) was applied to RD to obtain additional variables to map seagrass species and their AGCs (Wicaksono, 2016). Inversed Model (IM) water-column correction algorithm was used to minimize the effect of the water column on the reflectance of submerged seagrass (Wicaksono, Hafizt, 2013). The IM algorithm requires information on the depth of each pixel or bathymetry (z), the water-column attenuation coefficient for each band (ki), and the reflectance of optically deep water (RWi). The formula is as follows: The ki for Cyan, Blue, Green, Yellow, Red, and Red-edge bands is 0.009 m -1 , 0.018 m -1 , 0.072 m -1 , 0.075 m -1 , 0.178 m -1 , and 0.381 m -1 , respectively (Bukata et al., 1995;Wicaksono, Hafizt, 2013). The RWi for the WV2 image, as adapted from Wicaksono (2016), is 0.002 (Cyan), 0.011 (Blue), 0.015 (Green), 0.005 (Yellow), 0.003 (Red), or 0.003 (Red-edge). The z was generated empirically by calibrating optically shallow water reflectance with in situ bathymetry data using the band ratio method. A detailed explanation of the procedure to obtain the said bathymetry map is published in Wicaksono (2016). The resulting bathymetry map has a standard error of estimate (SEE) of 0.17 m for 0-1 m, 0.19 m for 1-2 m, 0.24 m for 2-3 m, 0.32 m for 3-4 m, and 0.78 m for 4-7 m; hence, the maximum effective depth is 7 m (Wicaksono, 2016). Beyond this depth, the SEE increases exponentially (3 m for depth >7 m) and is, therefore, practically unusable for water column correction input.

Seagrass species and AGC mapping
Seagrass species were mapped to create a seagrass species mask for AGC mapping at the species level. However, mapping seagrass species individually is challenging, as already been pointed out in Phinn et al. (2008), where the accuracy of seagrass species mapping using per-pixel classification, even with hyperspectral data, is low (<30%). Consequently, the classification scheme for the seagrass species mapping was constructed based on life form variations (Table 1).
Each photo-quadrat sample was labeled based on this classification scheme. Then, the samples were grouped into two, for training the model and classification algorithm and for assessing the accuracy: 80 samples for the class Ea (48 for training the model and classification; 32 for accuracy assessment), 50 for EaThCr (31; 19), and 83 for ThCr (53; 30).  (Wicaksono, Hafizt, 2013).
The Support Vector Machine (SVM) was applied only to pixels classified as seagrass to exclude non-seagrass pixels from the classification process and performed on RD, RB, and PC bands. SVM is a machine-learning classification algorithm that produces high classification accuracy, especially when the data are not normally distributed, as is frequently encountered in benthic habitat mapping with a detailed classification scheme. As a result, SVM has successfully mapped underwater objects with higher accuracy (Zhang et al., 2013;Eugenio et al., 2015). This study used the Gaussian Radial Basis Function (RBF) kernel in the SVM algorithm.
AGCseagrass maps were derived with stepwise regression analyses. There were six predictors for empirical AGC modeling: (1) RD bands, (2) ratio of RD bands, (3) IM bands, (4) ratio of RB bands, (5) Principle Component (PC) bands, and (6) ratio of PC bands. The regression analysis was performed at community and species levels, and the resulting regression functions were used to convert seagrass reflectance values into AGCseagrass.
The mapping accuracies were assessed with two approaches: the confusion matrix (Congalton and Green, 2019) for seagrass species classification and the SEE for AGCseagrass prediction. The SEE was converted into percent using confidence interval analysis based on the lower and upper limit of the mean of the reference data (Wicaksono et al., 2011). This analysis produced a range of expected minimum and maximum accuracies of the AGCseagrass map.

PCv and AGCseagrass measurement:
The measurement results of the harvested seagrass samples are shown in Table 2. Ea generally had lower PCv than other species because this class was found in lower density with wide gaps in between the shoots. However, its AGC was higher than other species, mainly because of its bigger leaf size. These findings provide the fundamental justification for predicting AGCseagrass using PCv and mapping it at the species level.

Seagrass class
PCv ( Table 2. PCv and AGC measurement results for each class of the harvested seagrass (n is the number of samples).

PCv-AGCseagrass analysis results at the community level:
The PCv had a correlation coefficient (r) of 0.37 (Sig. 95%) and R 2 of 0.14 with AGCseagrass. The low correlation can be addressed to variations in life forms and growth types across the seagrass species. The resultant regression function for predicting AGCseagrass from PCv was AGCseagrass = 0.1028(PCv) + 1.449.

PCv-AGCseagrass analysis results at the species level:
Analyses at the species level revealed a stronger PCv-AGCseagrass correlation than the one measured at the community level. It results from the seagrass biophysical properties that naturally vary with species. Larger species, such as Ea, grow vertically within a water column frequently up to the water surface. For seagrass species with a life form similar to Ea, their PCv may underestimate the actual AGCseagrass, as indicated by the R 2 of the regression analyses between PCvEa and AGCEa, which was the lowest of all seagrass species (r = 0.81, R 2 = 0.66). For Ea, the AGCseagrass was estimated using the regression equation AGCEa = 0.3179(PCvEa) + 0.6295.
Using PCv for predicting the AGCseagrass of horizontally growing species, such as Th, Cr, Hu, or Si, was less of an issue since an increase in PCv corresponded to an addition in the AGCseagrass. As a result, the R 2 between the PCv and AGCseagrass of these species was higher than Ea's.

Seagrass species composition mapping
The most accurate seagrass species composition map based on life forms had an OA of 76.11% obtained from PC bands. The ThCr classification was the most accurate, as shown by its high user's accuracy (UA) and producer's accuracy (PA) (  Table 3. Accuracy assessment results of the seagrass species composition mapping using SVM.

AGCseagrass mapping
The PCv interpreted from each photo-quadrat sample was used to estimate the AGCseagrass at the species level using the corresponding species composition class regression function. The AGCseagrass of each quadrat is the sum of the AGCs of all seagrass species present in the quadrat. Afterward, each sample was labeled based on the species composition class described in Table 1. The mean AGCseagrass from all photo-quadrat samples were 11.33 g C m -2 for training areas and 10.65 g C m -2 for validation samples.

AGCseagrass mapping at the community level:
The AGCseagrass was modeled and mapped at higher accuracy using band ratio (Table 4). The AGCseagrass map generated with the ratio of RD bands had 58.79% accuracy (SEE = 5.41 g C m -2 ), which was 10% higher than the best result of the single band model (RB bands, 48.69% accuracy, SEE = 6.69 g C m -2 ). The single band and band ratio results showed a similar pattern where AGCs below 10 g C m -2 were overestimated, while AGCs above 15 g C m -2 were underestimated. Thus, only AGCs in the range of 10-15 g C m -2 were correctly predicted. Table 4. Accuracy assessment results of the AGCseagrass mapping.

AGCseagrass mapping at the species level:
Compared with mapping at the community level (Table 4), the AGCseagrass maps generated at the species level were higher in accuracy for Ea and EaThCr but lower for ThCr. The mean AGCs were 15.41±6.07 g C m -2 for Ea (estimated from 46.67±19.21% PCv), 11.73±4.42 g C m -2 for EaThCr (64.49±17.22% PCv), and 4.89±2.16 g C m -2 for ThCr (50.91±19.82% PCv). Despite having the lowest PCv, Ea had the highest mean AGC because its species constituents had a significantly larger size than species in other classes.
The predicted AGCEa was saturated beyond 20 g C m -2 , while for the AGCEaThCr was saturated after 15 g C m -2 ; hence, the actual amount of aboveground carbon stored in seagrass with high predicted AGC is underestimated. In contrast, the predicted AGCThCr was overestimated, especially for AGC less than 8 g C m -2 . Based on these results, seagrass reflectance captured by remote sensing can correctly map AGCseagrass when it is in the range of 15-20 g C m -2 for Ea, 10-15 g C m -2 for EaThCr, and 4-8 g C m -2 for ThCr. Accordingly, the AGCseagrass in the study area was estimated at around 13.39 t C from 1.21 km 2 of seagrass habitat. In detail, AGCEa was estimated at around 0.35 t C, AGCEaThCr at 5.84 t C, and AGCThCr at 3.24 t C.

DISCUSSION
This research utilized seagrass PCv for rapid and non-destructive estimation to obtain AGCseagrass at community and species levels.
PCv and AGCseagrass show a weak correlation at the community level, causing the approximated AGCseagrass to be less accurate. Therefore, the equation provided in Wahyudi et al. (2020) is suggested for converting seagrass PCv into AGCseagrass without considering species variations. Nevertheless, the correlation coefficients and R 2 values are significantly higher when approximating AGCseagrass from PCv at the species level. Besides, estimating AGCseagrass from the PCv interpreted from photoquadrat samples is relatively fast and, thus, allows more samples to be collected.
At the community level, the highest AGCseagrass map accuracy is 58.79% (SEE = 5.41 g C m -2 ). To obtain better accuracy, mapping AGC at the species level is deemed necessary. It requires an accurate seagrass species map, and this research has successfully generated a seagrass species composition map with 76.11% OA. However, although this percentage is comparable to previous works by Roelfsema et al. (2014) and Koedsin et al. (2016), the accuracy of classes comprising individual seagrass species is not similarly high. In this research, Ea has a very low PA and underestimated extent. Under these conditions, it is also difficult to compare the AGCseagrass map accuracy achieved in this study with that in previous work. Therefore, the results of this research cannot be directly compared with other studies but can be used as the baseline for future mapping activities using seagrass classification with similar complexity.
This research also highlights some improvements in estimating and mapping the AGCseagrass at the species level. For instance, the AGCEaThCr map has the highest accuracy, followed by AGCEa and AGCThCr maps. Consisting of seagrass species with various life forms, EaThCr is expected to be the most challenging class to model. However, because its AGCs are in the range of 5 to 25 g C m -2 , species belonging to this class can be mapped at higher accuracy. Furthermore, Ea can be modeled quite accurately because the pixels categorized into this class are composed of only one species, with biomass having a negative linear correlation with the percent cover of sand or any background substrates. Therefore, the fraction of seagrass reflectance that composes the pixel mainly comes from Ea, and it effectively represents the abundance of Ea within the pixel. Assuming the background reflectance does not vary greatly, variations in biophysical properties are the main cause of changes in Ea reflectance. Moreover, Ea grows sparsely; thus, unless mixed with other species, bare substrates in between the shoots create larger gaps and produce background reflectance that affects the total reflectance of seagrass. This condition may also negate the image pixel value that is easily saturated. Consequently, Ea has a higher saturation threshold (20 g C m -2 ) than other classes (15 g C m -2 ).
Three important findings regarding the AGCseagrass mapping are: (1) mapping low AGCseagrass is difficult because the background reflectance largely contributes to and alters the resultant seagrass reflectance, (2) mapping high AGCseagrass is also challenging because seagrass reflectance is saturated at higher AGC due to the high energy absorption by seagrass bed with higher biomass, and (3) mapping seagrass with medium AGC is more feasible since the reflectance of background object is limited and yet to be saturated. This research acknowledges that one of the issues affecting the AGCseagrass mapping accuracy is the seasonal variations in seagrass presence, especially when the image acquisition does not temporally coincide with field data collection. This issue can be addressed in future studies by using a high temporal-resolution image, such as PlanetScope. Although PlanetScope has lower spectral and spatial resolution than WV2, it offers flexibility in selecting the acquisition date, thus enabling users to obtain images during field data collection and removing the error caused by seasonal variations in seagrass presence in the empirical AGCseagrass model.
Finally, the regression equations developed in this study are expected to improve the availability of AGCseagrass data that is currently lacking due to the laboratory analysis and destructive harvest involved in their acquisition, especially for seagrass species presented in this study. With more availability, this information can be used to train remote sensing images and validate the results of remote sensing-based AGCseagrass mapping.

Figure 2.
Life form-based seagrass species composition map with 76.11% OA, based on SVM classification results using WV2 PC bands (left) and AGCseagrass mapping at the community level with 5.41 g C m -2 SEE and 58.79% accuracy (right).

CONCLUSIONS
This research has developed a rapid aboveground seagrass carbon stocks (AGCseagrass) estimation using an easily measured parameter, seagrass PCv. Estimating AGCseagrass from PCv produces more accurate results when conducted individually for each species. Hence, this research recommends using speciesspecific PCv-AGCseagrass equation to approximate AGCseagrass from PCv. Furthermore, AGCseagrass can be modeled and mapped from remote sensing data. At the community level, the AGCseagrass map generated from WV2 images is 58.79% accurate (SE = 5.41 g C m -2 ). Meanwhile, at the species level, the accuracy increases significantly for Ea (64.73%, SEE = 6.86 g C m -2 ) and EaThCr (70.02%, SEE = 4.32 g C m -2 ) but decreases for ThCr (55.08%, SEE = 2.55 g C m -2 ). The results indicate that WV2 image reflectance can correctly map AGCseagrass in the study area when it is in the range of 15-20 g C m -2 for Ea, 10-15 g C m -2 for EaThCr, and 4-8 g C m -2 for ThCr. Finally, AGCseagrass in the study area is estimated at 13.39 t C.