Abstract
Wetland ecosystems play key roles in global biogeochemical cycling, but their spatial extent and connectivity is often not well known. Here, we describe an approach suitable for application on a 1000 km scale using Sentinel-1 and Sentinel-2 imagery, exploiting the implementation of Random Forest algorithm in Google Earth Engine. The approach was used to detect the spatial coverage of wetland types focusing on the case of southern Nigeria, thought to be one of the most wetland-rich areas of Africa. We compiled 1050 wetland and non-wetland control points for algorithm training and validation, primarily from visual interpretation of high-resolution (<1 m pixel) imagery. This allowed us to establish the relative importance of 18 input channels derived from Sentinel-1 polarimetric and Sentinel-2 indices for classification of wetland. We estimate that the swamps, marshes, mangroves, and shallow water wetlands of southern Nigeria cover 29,900 km² with 2% uncertainty of 460 km². We found larger mangrove and smaller marsh extent than suggested by earlier, coarser spatial resolution studies. Average continuous wetland patch areas were 120 km², 11 km², 55 km² and 13 km² for mangrove, marsh, swamp, and shallow water respectively. Our final map with 10 m pixels also captures small patches of wetland, with 20% of wetland patches being <1 km2; these were clustered around urban centres, suggesting anthropogenic wetland fragmentation. Our approach can now be used across rest of Africa and globally to detect wetlands and wetland change which, in turn, will be crucial for improved land-surface climate models and wetland conservation.