Wetlands Ecology and Management Ecology, Evolution, Behavior and Systematics

Improving and evaluating local-scale wetland representation using multi-source remote sensing data and object-based machine learning in the Kuils River Catchment, South Africa

Gxokwe, Siyamthanda, Dube, Timothy, Ngcoliso, Nasiphi, Ngebe, Sisipho
Springer Nature · 2026-11-21 · DOI 10.1007/s11273-026-10215-y

Abstract

Wetlands are vital ecosystems that support biodiversity, regulate key hydrological processes (including floodwater attenuation, aquifer recharge, and baseflow maintenance), enhance water quality through nutrient transformation and contaminant retention, and serve as vital biogeochemical carbon sinks. However, urban expansion, agricultural activities, and limitations in existing wetland inventories threaten their conservation and accurate representation. This study evaluated wetland representation in the Kuils River Catchment, South Africa, using multisource Earth Observation (EO) data and geospatial analysis. Sentinel-1 Synthetic Aperture Radar (SAR), Sentinel-2 optical imagery, and Global Ecosystem Dynamics Investigation (GEDI)-derived vegetation structure data were integrated within an object-based machine learning framework to produce a high-resolution wetland extent layer. The resulting map was compared with existing datasets, including the South African National Space Agency Mzansi Amanzi dataset, the Global Surface Water product, and the South African National Land Cover 2022 layer. The results indicated that floodplain wetland systems were prominent within the catchment. The classification achieved an overall accuracy exceeding 90%, although class-specific performance and remaining uncertainties should be considered when interpreting the results. Comparisons with existing products identified differences in wetland representation, with some small and potentially seasonal wetland areas less extensively represented in broader-scale datasets. The integration of radar, optical, and vegetation structure information provided complementary data for distinguishing wetlands from non-wetland areas. Nevertheless, uncertainties associated with surface moisture sensitivity and differences in the spatial resolution of GEDI and Sentinel datasets may have influenced classification outcomes. The findings demonstrate the potential of multisource EO data and object-based machine learning to support local-scale wetland mapping. The resulting wetland extent layer may inform conservation, land-use planning, and environmental management within the catchment. Further validation across seasons and other catchments is needed to assess the transferability of the approach.

Read at publisher →