Abstract:
Wetlands are recognised as the important natural ecosystems in the world. The above-ground
biomass (AGB) of wetland vegetation is essential for providing ecosystem services related to
global climate change due to its crucial role in sequestrating anthropogenic carbon emissions.
Seasonal AGB estimation could help to understand carbon changes in wetlands and how
vegetation in these ecosystems differs across seasons at regional scales. Remote sensing
technology offers time-effective and cost-efficient ways to improve the monitoring of wetlands
and understanding of the spatial carbon changes in wetland vegetation. This study aimed to use
seasonal derived AGB of palustrine herbaceous vegetation to determine the differences in teal
carbon, using active and passive remote sensing data across the summer and winter seasons. The
study was carried out in the Chrissiesmeer catchment in the temperate Grassland Biome of the
Mpumalanga Province of South Africa. The objectives were to (1) derive different season-specific
modelling scenarios from Sentinel-1 and Sentinel-2 imagery to assess the optimal model for
estimating AGB of palustrine wetland vegetation AGB, (2) assess the performance of Random
Forest (RF) and Support Vector Regression (SVR) in predicting seasonal AGB of wetland
vegetation, (3) map the seasonal spatial patterns of teal carbon from the estimated AGB of
wetland vegetation, and (4) assess the seasonal variation in the predicted teal carbon. RF and
SVR algorithms were used as regression-based algorithms with important variable selection to
develop an optimal model from the modelling scenarios, which also incorporated field-measured
Leaf Area Index (LAI). The results showed that the combination of Sentinel-1 GLCMs and
backscatter channels yielded higher accuracy for the estimation of the AGB of palustrine
herbaceous vegetation attaining coefficient of determination (R2
) = 0.735, root mean squared
error (RMSE) = 39.848 g·m-2
, and relative RMSE (relRMSE) = 17.286% compared to a combination
of reflectance bands, vegetation indices and red-edge bands (R2
= 0.753, RMSE = 49.268 g·m-2
,
and relRMSE = 20.009%) in the summer season. For the estimation of AGB in the winter season,
Sentinel-1-derived GLCMS textures obtained higher accuracy (R2 = 0.785, RMSE = 67.582 g·m-2
,
and relRMSE = 20.885%) compared to the combination of reflectance bands, vegetation indices
and red-edge bands of optical data (R2 = 0.749, RMSE= 69.634 g·m-2 and relRMSE = 21.248%).
xv
These findings suggested that Sentinel-1 sensor-derived models performed better than the
optical models in both seasons. Furthermore, the addition of SAR textural measurements
improved the accuracy of modelling AGB and RF model performed better than SVR in estimating
the AGB of wetland vegetation. The study observed that there was a significant difference
between the summer (77.527 g C/m-2 DM) and winter (57.918 g C/m-2 DM) seasonal mean carbon
ranges (p < 0.05), and Tevredenpan wetland vegetation communities stored higher levels of
carbon in the AGB vegetation in summer than in winter. The study showed that vegetation of
palustrine wetlands is significant for carbon storage and fluctuates significantly between summer
and winter. Estimating carbon stock in the AGB vegetation can aid in conserving grasslands and
wetlands and notably optimise research on biomass estimation with remote sensing and machine
learning systems.
