Functional and variables selection in extreme value models for regional flood frequency analysis

TL;DR

This paper introduces a Bayesian P-spline based approach with grouped horseshoe priors for functional and variable selection in GEV models, improving flood frequency analysis by reducing prediction uncertainty.

Contribution

It presents a novel Bayesian modeling framework that incorporates non-linear covariate effects and variable selection for regional flood frequency analysis.

Findings

The proposed model reduces uncertainty in ungauged flood predictions.

It outperforms alternative models in cross-validation studies.

The approach effectively identifies relevant covariates and non-linear effects.

Abstract

The problem of estimating return levels of river discharge, relevant in flood frequency analysis, is tackled by relying on the extreme value theory. The Generalized Extreme Value (GEV) distribution is assumed to model annual maxima values of river discharge registered at multiple gauging stations belonging to the same river basin. The specific features of the data from the Upper Danube basin drive the definition of the proposed statistical model. Firstly, Bayesian P-splines are considered to account for the non-linear effects of station-specific covariates on the GEV parameters. Secondly, the problem of functional and variable selection is addressed by imposing a grouped horseshoe prior on the coefficients, to encourage the shrinkage of non-relevant components to zero. A cross-validation study is organized to compare the proposed modeling solution to other models, showing its potential in reducing the uncertainty of the ungauged predictions without affecting their calibration.