Abstract:
Monitoring changes in the areal extent and geographic distribution of wetland vegetation has become more critical considering the impact of anthropogenic and climate changes. We compared the capabilities of the optical space-borne sensors Sentinel-2 and WorldView-3 (WV3) to distinguish between wetland and terrestrial vegetation for improved reporting to the Sustainable Development Goal (SDG) sub-indicator 6.6.1a, and also map different wetland vegetation communities for two catchments in the Grassland Biome of South Africa. Ground truthing of vegetation communities was conducted between 2016 and 2018. A Random Forest classification algorithm was used with a 100-fold cross-validation to assess mean accuracies using all combinations of bands, a digital elevation model generated from fine-scale contours, spectral vegetation indices (VIs) and above-ground biomass (AGB). Five and eight wetland vegetation classes were mapped for Hogsback and Tevredenpan, respectively, of a total of 13 classes for each of the sites. Wetland and terrestrial vegetation were found to be highly separable, with overall accuracies (OAs) attaining 91–99% and individual user's accuracies 88–99% for both sensors and study areas. Even though the wetland vegetation communities consisted of a mosaic of smaller communities, monodominant species and plant functional type classes, they were found to be highly separable across sensors and study areas. The highest average OA of 83% for Hogsback's wetland vegetation communities was achieved using WV3 bands with elevation, AGB and the VIs, while the Sentinel-2 bands, elevation, AGB and VIs attained an average OA of 78%. For Tevredenpan, the use of the Sentinel-2 bands and elevation achieved the highest mean OA of 79% for the classification of wetland vegetation communities, while the WV3 (in this case the short-wave infrared bands were not available owing to shortage of funding) maximized at 74%. The inclusion of elevation data and spectral indices in the classification scenarios of wetland vegetation communities increased the OA by 4–17%. Omitting the red-edge and shortwave infrared bands for classification of vegetation classes resulted in a varied response across sensors and study areas, but decreased the OA by 4.8–7.3% when using the Sentinel-2 sensors. These results show promise for improved reporting and monitoring of the extent and types of palustrine wetlands in the Grassland Biome of South Africa using freely-available Sentinel-2 data.
