Quantifying uncertainty in deterministic observed meteorological datasets: A case study applied to large-domain gridded NLDAS-2 daily precipitation and temperature fields

Abstract:

Land surface and hydrologic modeling applications are characterized by large uncertainty in the observed meteorological forcing data, yet few ensemble historical meteorological datasets are available, and most existing large-domain meteorological surface observation datasets are deterministic. There is a need to provide better tools and information to support users in augmenting such deterministic meteorological datasets to yield uncertainty estimates that can be applied in applications such as forecasting and data assimilation. This study demonstrates the use of a locally-weighted spatial regression method to quantify the uncertainty in precipitation and temperature fields in gridded historical meteorological datasets. The method uses spatial attributes from sampled neighboring grids as the explanatory variables to estimate the variability of precipitation and temperature over the spatial fields, with the regression uncertainty serving as a basis for generating an ensemble of meteorological variables. We demonstrate this approach through a CONUS-wide application to the widely used 1/8th degree NLDAS-2 precipitation and temperature dataset. This study assesses the viability of the approach through comparison of the resulting precipitation and temperature uncertainty estimates to those obtained from the technique when applied to raw station observations. We also illustrate the application of the approach through case studies in which meteorological uncertainty is propagated to uncertainty in ensemble hydrologic simulations.