A multiple lines of evidence approach for choosing at-site nonstationary flood-frequency analysis methods

Abstract:

Although research on nonstationary flood frequency analysis (NSFFA) has proliferated, few continental-scale studies have compared the performance of NSFFA methods for updating design flood events to reflect current conditions. Moreover, practitioners have little guidance for considering the inherent biases and uncertainties of these methods along with their fit to observed annual peak flows. First, to compare the inherent biases and uncertainties of six NSFFA methods, we parameterize a Monte Carlo experiment using distribution properties and trends observed in annual peak flow series in the conterminous United States. NSFFA methods were selected to examine trends in both central tendency and variability and included approaches based on Ordinary Least Squares regression (OLS), Gamma Generalized Linear Models (GLMs), the Generalized Additive Model for Location Scale and Shape (GAMLSS), and quantile regression, a distribution-free approach.

Experiment results are illustrated using trend-space plots that indicate trend magnitudes above which modeling changes in the central tendency and variability is warranted using different model performance criteria. We examine the effects that distribution properties, such as skewness, have on these trend thresholds and NSFFA method choices. For instance, quantile regression performs much better than distribution-based methods when log-transformed peak flows are negatively skewed while GLM or GAMLSS are often preferable when skewness is positive. We also introduce an approach for comparing the goodness-of-fit of design floods estimated from assumed theoretical distributions and quantile regression. Through case studies of watersheds experiencing pronounced multi-decadal climate variability and urbanization, we demonstrate the extent to which Monte Carlo experiments and goodness-of-fit analyses can, together, inform NSFFA method selection.