Resampling Gaged Networks to Provide Uncertainty Estimates for Daily Streamflow Predictions in Ungaged Basins

Abstract:

In ungaged basins, predictions of daily streamflow are essential to responsible and effective management and design of water resources systems. Transfer-based methods are widely used for prediction in ungaged basins (PUB) within a gaged network. Such methods rely on the transfer of information from an index gage to an ungaged site. In what is known as the nearest-neighbor algorithm, the index gage is selected based on geospatial proximity. The predictions offered by any PUB method can be highly uncertain, and it is often difficult to characterize this uncertainty. In the development of predicted streamflow records, understanding the uncertainty of estimates would greatly improve water resources management in ungaged basins. It is proposed that by resampling the sites of the gaged network, with replacement, a set of equally-probable streamflow predictions can be produced for any ungaged site. For a particular day in the record, the percentiles of the distribution of the resampled, predicted streamflows can be used to estimate confidence intervals of the original daily streamflow predictions. This approach is explored in the Southeast United States with a nearest-neighbor application of non-linear spatial interpolation using flow duration curves (QPPQ), a common PUB method. Though some interval re-centering is required to ensure that the best-case prediction falls within the confidence intervals, it is shown that this technique provides a reasonable first-order approximation of prediction uncertainty. Still, the best estimated confidence intervals are shown to consistently under-estimate the nominal confidence. It is hypothesized that this interval contraction is a result of temporal and spatial correlation within the gaged network. Additionally, implications of prediction uncertainty are explored and alternative estimators are considered.