Development of a global ~90 m water body map using multi-temporal Landsat images

This paper describes the development of a Global 3 arc-second Water Body Map (G3WBM), using an automated algorithm to process multi-temporal Landsat images from the Global Land Survey (GLS) database. We used 33,890 scenes from 4 GLS epochs in order to delineate a seamless water body map, without cloud and ice/snow gaps. Permanent water bodies were distinguished from temporal water-covered areas by calculating the frequency of water body existence from overlapping, multi-temporal, Landsat scenes. By analyzing the frequency of water body existence at 3 arc-second resolution, the G3WBM separates river channels and floodplains more clearly than previous studies. This suggests that the use of multi-temporal images is as important as analysis at a higher resolution for global water body mapping. The global totals of delineated permanent water body area and temporal water-covered area are 3.25 and 0.49 million km2 respectively, which highlights the importance of river-floodplain separation using multi-temporal images. The accuracy of the water body classification was validated in Hokkaido (Japan) and in the contiguous United States using an existing water body databases. There was almost no commission error, and about 70% of lakes > 1 km2 shows relative water area error < 25%. Though smaller water bodies (< 1 km2) were underestimated mainly due to omission of shoreline pixels, the overall accuracy of the G3WBM should be adequate for larger scale research in hydrology, biogeochemistry, and climate systems and importantly includes a quantification of the temporal nature of global water bodies