COMBINING REGIONAL REGRESSION APPROACHES WITH GEOSTATISTICAL TECHNIQUES FOR PREDICTING FLOOD QUANTILES IN UNGAUGED BASINS

The study explores the application of two different spatial interpolation techniques, Generalized Least Squares (GLS) and Top-kriging, aiming to enhance the prediction performances by blending these two powerful methods. GLS is one of the most common approaches used in United States (US) for predicting flood quantiles in ungauged basins, Top-kriging predicts streamflow statistics along river networks (e.g. flood quantiles) taking both the catchment area and nested nature of catchments into account. We first applied GLS in cross-validation, through a comprehensive leave-one-out procedure, to predict flood quantiles over a wide region in the southeast of US by simulating the ungauged conditions at each and every site, then we used Top-kriging to model GLS residuals. The results reveal that combining GLS with Top-kriging increases the prediction capability of GLS. Blending the techniques increases the Nash-Sutcliffe efficiency from 0.67-0.76 to 0.75-0.91, depending on the frequency of the quantile; moreover blending GLS with Top-kriging reduces the prediction errors in the majority of the study catchments relative to using GLS alone.