Determining Uncertainty in Estimations of Global Surface Water Extent derived from a Diurnal Earth Observation Time-Series

Abstract

The availability of fresh water is vital for life on the planet. However, water resources are increasingly affected by changing patterns of climate variables and intensification of human water use and regulatory management. Consequently, focus of many disciplines is directed towards inland water storage dynamics. Satellite remote sensing offers opportunities to monitor global surface water in dense temporal intervals. The DLR-DFD Global WaterPack (GWP) provides daily information on inland surface water on a global level. The dataset has been successfully applied in manifold scientific studies, enabling the investigation of surface water dynamics world-wide. To enhance the usability of the product towards modeling applications, the quantification of inherent uncertainties is essential. As GWP is primarily based on optical MODIS (Moderate Resolution Imaging Spectroradiometer) images, inaccuracies arise due to coarse spatial resolution and interpolation of data gaps. To address these error sources, we quantify interpolation- and classification-based uncertainty in two steps. First, several spatiotemporal time-series characteristics relevant for water body dynamics are considered to determine the probability of interpolated pixels to be covered by water. Second, in case of valid observations, GWP classification probability is derived from relative datapoint (pixel) locations in feature space and subsequently utilized together with previously established temporal information in a linear mixture model. Resulting sub-pixel water fraction estimates facilitate the quantification of observational uncertainty. Performance of water fraction estimates is assessed in 32 regions of interest across the globe by comparison to higher resolution reference data. The ability of temporal layers to approximate unknown pixel states is evaluated for artificial gaps, introduced to the original time-series of four MODIS tiles. Results show that uncertainties can be quantified accurately, revealing more comprehensive and reliable time-series information suitable for modelling applications.