Assessment and attribution of predictive hydrologic uncertainty in operational short-to-medium range ensemble streamflow forecasting

We assess the predictive hydrologic uncertainty in short-to-medium range (< Day 14) ensemble streamflow forecasting via 20-yr hindcast experiments for 398 headwater locations in the service areas of the 13 River Forecast Centers in the US. The assessment is based on the continuous ranked probability skill score (CRPSS) of ensemble streamflow prediction under clairvoyant future input forcings in reference to climatological ensemble streamflow forecast. The magnitude of hydrologic uncertainty is assessed in the CRPSS space relative to the total uncertainty represented by climatological ensemble streamflow forecast. Hydrologic uncertainty is attributed to selected uncertainty sources by isolating the lead time-dependent initial condition (IC) uncertainty via state updating. The limits of predictive skill under the current state of operational hydrologic modeling are assessed via the CRPSS of perfectly reliable ensemble streamflow predictions with and without state updating. The results show that there exists large room for improvement in forecast accuracy from reducing and accounting for hydrologic uncertainty. Significant unresolved hydrologic uncertainty exists for almost all locations for prediction of high flows. For prediction of all ranges of flow, unresolved hydrologic uncertainty is larger in the CRPSS space than the uncertainty resolved by hydrologic modeling and clairvoyant future forcings for most locations. The IC uncertainty at saturation is very significant and accounts for more than half of the unresolved hydrologic uncertainty on average. Isolation of the IC uncertainty indicates that many locations will likely benefit significantly from additional calibration of hydrologic models and fully utilizing the hydrologic memory via high-frequency state updating.