Prediction of regional flow duration curves: geostatistical techniques versus multivariate regression

Wednesday, 16 December 2015
Poster Hall (Moscone South)
Alessio Pugliese1, William Hastings Farmer2, Attilio Castellarin1, Stacey A Archfield3 and Richard M Vogel4, (1)University of Bologna, Bologna, Italy, (2)USGS Colorado Water Science Center Denver, Denver, CO, United States, (3)USGS Groundwater Information, Reston, VA, United States, (4)Tufts University, Department of Civil and Environmental Engineering, Medford, MA, United States
A period-of-record flow duration curve (FDC) represents the relationship between the magnitude and frequency of daily streamflows. Prediction of FDCs in ungauged basins is of great importance in those locations characterized by sparse, or more often missing, streamflow observations. We present a detailed comparison of two approaches which are capable of predicting an FDC in ungauged basins. An adaptation of the geostatistical method Top-kriging employs a linear weighted average of dimensionless empirical FDCs, standardized for a reference streamflow value. Weights are the result of the application of Top-kriging over a point index which, empirically, expresses the similarity between curves. Dimensional FDCs are then reconstructed developing a similar Top-kriging-based model capable of predicting the reference streamflow in the same sites. The second method is based on regional multiple linear regressions and is one of the most common method for prediction of FDCs in ungauged sites. Comparisons of these two methods are made at 182, mostly unregulated, river catchments in the southeastern U.S. using a three-fold cross-validation algorithm. Our results reveal that the two methods perform very similarly throughout flow-regimes, showing average Nash-Sutcliffe Efficiencies of 0.566 and 0.662 in natural scale, while 0.883 and 0.829 in log-transformed scale, for the geostatistical and the linear regression models, respectively. However, some complementarities are shown in the very low-flow regime, i.e. duration greater than 0.95, where the two models highlight different behaviors whether considering natural or log-transformed streamflows.