Seasonal Mapping of Urban Surface Water Bodies Using Sentinel-2 Imagery and Random Forest Classification: A Case Study of Mbombela, South Africa

Urban surface water bodies regulate hydrological processes, support biodiversity and provide municipal and ecological services in rapidly expanding cities. Small and fragmented water features in urban and peri-urban landscapes remain difficult to map at medium resolution because of mixed pixels, spectral confusion between water, shadows and dark impervious surfaces, and strong seasonal variability. This study mapped seasonal urban and peri-urban surface water in Mbombela, South Africa, from Sentinel-2 observations acquired between 2024 and 2026 using Random Forest classification.

Sentinel-2 Surface Reflectance scenes spanning 2024–2026 were processed in Google Earth Engine. Cloud-masked wet-season and dry-season composites were generated for each year in the three-year period. Predictor variables comprised visible, near-infrared and shortwave-infrared bands together with NDWI, MNDWI, AWEI, NDVI and NDBI. A Random Forest classifier was trained to separate water from vegetation, built-up land, bare soil and agricultural land. Accuracy was evaluated with independent validation samples interpreted from high-resolution Google Earth imagery and field observations. Seasonal water extent was compared across the 2024–2026 composites.

Independent validation produced overall accuracies ranging from 74% to 79% for the seasonal classifications derived from the 2024–2026 Sentinel-2 record. Performance varied with seasonal hydrology and residual spectral confusion at water–built-up and water–vegetation edges, including shadowed urban surfaces. Across 2024–2026, the mapped water extent was larger and more spatially connected in wet-season composites than in dry-season composites, when many small peri-urban features contracted or disappeared. The multi-index Random Forest models reduced commission errors relative to single-threshold NDWI or MNDWI masks, although omission of narrow channels persisted at 10 m resolution.

The 2024–2026 Sentinel-2 time series and Random Forest classification provide a reproducible seasonal inventory of urban and peri-urban surface water in Mbombela. Overall accuracies of 74–79% indicate operational utility for municipal hydrological planning over this recent period, while underscoring the need for finer-resolution or multi-sensor inputs where small, fragmented water bodies dominate.