Assessment of WorldView-2 images for aboveground seagrass carbon stock mapping in patchy and continuous seagrass meadows

Spatially explicit information on aboveground seagrass carbon stock (AGC(seagrass)) is required to understand the role of seagrass as a nature-based solution to mitigating and adapting to climate change. Remote sensing provides the most effective and efficient means to map AGC(seagrass). This research aimed to assess the accuracy of multispectral images to map AGC in seagrass meadows with different characteristics. Kemujan Island in Indonesia was selected to represent patchy beds with relatively low seagrass percent cover and low water clarity, whereas Labuan Bajo of Indonesia served an example of continuous beds with a relatively high percent cover and water clarity. WorldView-2 is a high-spatial-resolution multispectral image with six visible bands, by which water can be penetrated. Several image processing techniques, including sunglint and water column correction, principal component analysis, and co-occurrence texture analysis, were applied to the atmospherically corrected WorldView-2 images. These datasets were used as inputs in the AGC empirical model using seven regression algorithms, namely, single-band linear regression, band-ratio linear regression, stepwise regression, random forest regression, support vector regression, extreme gradient boosting, and multivariate adaptive regression spline. Seagrass field data collected using photo quadrat technique were used to train the regression model and assess the accuracy of the resulting AGC(seagrass) maps. The results indicated that WorldView-2 can be used to map AGC in different seagrass meadows with a consistent accuracy. The most accurate AGC(seagrass) map for Kemujan Island had a root mean square error (RMSE) of 4.11 g C m(-2) for the aboveground stock in the range of 0-28.70 g C m(-2), and for Labuan Bajo, the RMSE of the map was 9.73 g C m(-2), with aboveground stock range of 0.31-64.38 g C m(-2). Model cross-validation revealed that the mapping model can be site-specific or robust depending on the characteristics of the field-derived AGC(seagrass) data used to train the regression algorithm. For example, the model developed for Labuan Bajo seagrasses, which had a higher AGC variance, can be successfully applied to Kemujan Island with its lower AGC variance, but not vice versa. This finding is a key factor in the future development of a robust AGC(seagrass) mapping model that is applicable to various seagrass meadows as a stepping stone to accelerating the filling of gaps in the global seagrass dataset.