An exploration of post-processing uncertainty estimation in conceptual rainfall-runoff modelling
Year: 2026
Type: dissertation
Abstract: Conceptual rainfall-runoff models are widely used in hydrology and represent a primary tool for generating predictions, particularly in data-scarce contexts.However, in many practical applications, these models are still employed in a deterministic framework, without adequately accounting for the multiple sources of uncertainty affecting their outputs.This thesis provides a comprehensive investigation of uncertainty in conceptual rainfall-runoff modelling, with the dual objective of clarifying its main sources and developing practical tools for its quantification in operational settings.Chapter 2 begins with the development of a large-scale regional rainfall-runoff model over the Greater Alpine Region, with a focus on reconstructing flood events.This application serves as an example to highlight key modelling limitations, showing how uncertainties increase significantly near extreme events and when models are applied to ungauged catchments.These findings demonstrate that neglecting uncertainty can lead to misleading results, particularly in operational contexts.Building on this evidence, Chapter 3 investigates the statistical properties of rainfall-runoff model residuals through a large multi-catchment, multi-model study across the continental United States.The analysis focuses on key features such as marginal distributions, dependence structures, and temporal correlations.Both general patterns and catchment-specific behaviors are identified and related to climatic and morphological characteristics.In addition, the impact of residual transformations aimed at stabilizing statistical properties is assessed, highlighting which features can be effectively controlled and which remain inherently difficult to model within standard error modelling frameworks.Based on these insights, Chapter 4 then explores practical post-processing approaches for aggregated uncertainty estimation.The focus is placed on parametric methods, due to their potential improvements in out-of-sample applications and regionalization.A first approach consists of a copula-based method (C-method) designed to stochastically generate residual and discharge time series.Its performance is evaluated against established non-parametric resampling techniques, showing comparable performance but also revealing limitations that would require increased model complexity, potentially hindering regional applicability.To address this issue, in Chapter 5, a second approach is developed, focusing on the direct estimation of confidence intervals without generating full time series.This method, formulated as a parametric copula-based reinterpretation of the BLUECAT (Koutsoyiannis and Montanari, 2022) framework, is implemented in two versions targeting different flow regimes.The introduction of parametric tail modeling allows extrapolation beyond observed extremes.Results indicate performance comparable to non-parametric methods, with improved estimation of uncertainty for high-flow conditions, as confirmed by peak-over-threshold analyses.Chapter 6 of the thesis addresses uncertainty estimation for predictions in ungauged catchments through a regionalization of the previously proposed parametric BLUECAT framework.Model components, including deterministic predictions, copula dependence structures, and discharge distributions, are transferred to ungauged sites to estimate confidence intervals.The results are again comparable to those obtained with non-parametric approaches, while highlighting areas for methodological refinement.Overall, the findings do not support, within the tested conditions, the hypothesis that parametric post-processing methods systematically outperform simpler nonparametric alternatives.At the same time, the results identify clear directions for future research, suggesting that further methodological developments may enable parametric approaches to achieve improved performance, especially in ungauged catchments, in uncertainty estimation for rainfall-runoff modelling. PREFACEIt is always difficult to find the right way to begin a paper, or in this case, a Ph.D. thesis.One tends to search for the perfect opening sentence, something that can immediately set the stage and capture the spirit of what follows.Unfortunately, this has never been one of my strengths.I have always been more comfortable doing things than describing them in writing.For this reason, I hope the reader will allow me, in this opening paragraph (and only in this one), a slightly more informal and less pragmatic tone, while I briefly introduce the motivations behind the research questions explored in this work.Like most students embarking on a Ph.D., I began this journey with a general topic and a few core ideas already in mind.Still, it would have been impossible to foresee all the challenges, detours, and unexpected turns that would appear along the way, as well as the achievements that would eventually emerge from them.In fact, I should probably consider myself fortunate that, despite the different directions explored over the past three years, some more successfully than others, the topics I worked on always retained a common thread.From my perspective, this has allowed the final result to remain coherent, rather than becoming a simple collection of loosely connected pieces.Three main themes have guided the work presented in this thesis: uncertainty estimation, prediction in ungauged basins, and river floods.None of these are new problems in hydrology; many researchers before me have studied one or more of them in great depth.How this work attempts to contribute, however modestly, to these topics will form the core of the following chapters.Another element that deserves to be mentioned from the start is the main tool used throughout the research.Among the many possible approaches available to tackle these problems, we chose to rely on conceptual rainfall-runoff models.Finally, I would like to briefly highlight the general philosophy that guided the research here presented.Although a Ph.D. is often primarily expected to focus on fundamental research, we tried to keep in mind what this work might eventually enable, if not immediately, then in future developments.In other words, the aim was to avoid research carried out purely for its own sake, and instead to pursue ideas that might eventually prove useful in practical or operational contexts.This perspective influenced many of the choices made along the way.Often, when several alternatives were available, the option that appeared simpler, but potentially more attractive for real-world applications, was preferred.The idea of producing research that would remain confined to the pages of an article or this thesis, without ever potentially contributing to practical use, simply to say "it was done," was not particularly appealing to us.This is not meant as criticism of purely theoretical research; rather, it reflects my own background and inclination as an engineer.Whether this goal has actually been achieved is, of course, not for me to decide.But now that the ice has been broken, I will leave the reader to the more formal introduction that follows, hoping that the rest of this thesis will prove, at least to some extent, both interesting and enjoyable to read.
Source: PORTO@Iris
Author: LOMBARDO, LUCA
Institution: Politecnico di Torino
Language: English
Subfield: Water Science and Technology
Field: Environmental Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7218105797
