Hierarchical classification of urban LULC using Sentinel-2 and WorldView imagery: A comparative study for conservation applications

Accurate and operationally viable mapping of urban wetlands and land cover remains a persistent challenge for conservation authorities managing fragmented natural landscapes under increasing development pressure. Few studies have directly compared freely available medium-resolution imagery against costly very-high-resolution alternatives specifically within urban wetland contexts, leaving agencies without clear evidence-based guidance on sensor investment decisions. This study presents a two-stage hierarchical Random Forest classification framework implemented in Google Earth Engine (GEE), using multi-year seasonal Sentinel-2 (S2, 10 m) and Visible Near Infrared (VNIR) WorldView-2/3 (WV, 1.8 m and 1.2 m) imagery to map broad urban land cover and wetland subclasses within a 50.7 km 2 peri-urban conservation area in Brampton, Ontario, Canada. Each sensor was evaluated independently under identical algorithmic and training conditions. Multi-seasonal band stacking consistently outperformed single-season imagery for both sensors, improving overall accuracy by 5.1% and 8.6% for S2 and WV respectively, suggesting that, in this study context, temporal feature richness is a stronger driver of classification performance than spatial resolution alone. At the broad class level, S2 and WV achieved comparable overall accuracies of 91.2% and 90.5% respectively. S2 demonstrated consistent advantages of moisture-sensitive natural classes including wetland and herbaceous, attributed to its SWIR and red-edge bands which provide spectral sensitivity unavailable in the standard WV multispectral product. For wetland subclass mapping, S2 outperformed WV for bog and swamp classification while both sensors performed comparably for marsh. These findings demonstrate that for pixel-based Random Forest framework, freely available S2 imagery, implemented within a scalable cloud-based GEE workflow, matches or exceeds the classification performance of standard VNIR WV product for urban LULC and wetland monitoring, offering conservation organizations a practical and transferable approach for routine wetland inventory and long-term LULC monitoring.