Catchment similarity and spatial correlation: added value and impacts on hydrological predictions in ungauged basins

Abstract

This Thesis presents a three-fold study aimed at deepening our understanding on the added value and impacts of catchment similarity and spatial correlation (or cross-correlation) on the regional prediction of flood quantiles and flow-duration curves (FDCs) in ungauged river cross-sections. First, we consider the reference procedure for design flood estimation in Triveneto, North-eastern Italy, which assumes the entire study area to be a single hydrologically homogeneous region. We highlight that Triveneto cannot be regarded as homogeneous in terms of flood frequency regime and show that a focused-pooling approach accounting for selected geomorphoclimatic descriptors leads to regional samples with significantly improved homogeneity. Nevertheless, focused pooling does not consider the effects associated with cross-correlation, which are instead considered by Generalized Least Squares (GLS) and Top-kriging (TK; geostatistical method), although in two different ways. Recent studies show that TK outperforms GLS for predicting empirical flood quantiles, but they also speculate that cross-correlation might affect their accuracy in predicting true flood quantiles. To investigate this aspect, we apply GLS and TK for predicting flood quantiles in a homogeneous pooling-group of sites in Triveneto under different cross-correlation scenarios through a Monte Carlo experiment. For both methods, we observe that an increasing degree of spatial correlation results in an increasing masking-effect on the true flooding potential. Morever, we confirm that TK significantly outperforms GLS when they both assume flood quantiles to scale with drainage area alone, yet, both methodologies significantly improve their accuracy when considering several catchment descriptors. Finally, concerning the prediction of FDCs in a large and heterogeneous region, the Danube river basin, we show that geostatistical models provide much more accurate predictions than multi-regression models. In summary, all the analyses confirm the added value for statistical regionalisation of properly handling hydrological heterogeneity, also highlighting the pivotal role played by cross-correlation in observed streamflow time-series.