Spatial uncertainty to determine reliable daily precipitation maps

Abstract

Daily precipitation observations are commonly used with related variables to make estimates at unsampled locations to provide maps and gridded data for hydrological and climate model applications. Uncertainty in the way gridded data (maps) are prepared, given the available information, is rarely considered. Over a study period of one year, we used conditional simulations to produce multiple equally likely realizations of Australian precipitation per day. Together those realizations represented an ensemble measure of spatial uncertainty for rainfall for a given day. An independent gauge data set had values within the 5th–95th percentile uncertainty range 94% of the time. Combined with other measures they confirmed the reliability of the ensemble spatial uncertainty ranges. We compared several established mapping techniques to an independent gauge data set using local error statistics and to the spatial uncertainty maps. Those statistics showed little difference between the mapping techniques and overall assessment of performance was largely dependent on skill scores. However, the mapping techniques were different when compared to the spatial uncertainty ranges. These findings support the assertion that assessment of mapping techniques using local error statistics is insensitive to the uncertainty in producing the maps as a whole. We conclude that uncertainty information in precipitation estimates should not be overlooked when comparing precipitation estimation techniques. The focus of performance assessment is traditionally on local error estimates, and this tradition diverts attention away from the issues of uncertainty and reliability. Reliable uncertainty characterization is necessary for the rigorous detection of spatial patterns and longer time series trends in precipitation.