During the last decade, a series of large river ooding events (e.g. 2005 in the alpine region, 2013 in central Europe, or 2014 in northwestern Italy) have caused severe damages in Europe. An improved concept of integrated flood risk management in necessary, in order to manage and minimize flooding risks in the future, and to reduce the catastrophic nature of these events. A crucial step in any flood risk assessment is the accurate estimation of extreme ood peak discharges associated with a very low exceedance probability, i.e. with high return periods, which are then used for hydraulic design purposes and risk zone mapping. The uncertainties involved in these estimates need to be quanti ed, for reliable decision making in these tasks. The aim of this thesis is to better understand the sources and nature of dierent uncertainty components present in flood frequency estimation, and provide methods for quantifying them at dierent spatial and temporal scales. If discharge measurements are not available, flood design values can be computed either from precipitation data by rainfall-runo modelling, or by transferring the flood regime information to the target site from neighboring donor catchments with statistical regionalisation methods. In the context of the latter, Chapter 2 of this thesis investigates the uncertainties involved in predicting flood frequencies in ungauged catchments as a function of climate, method, and data availability. A global meta-analysis of the existing literature in the last twenty years is performed, involving a total of 3023 catchments worldwide. The reported cross-validation predictive performances of regionalisation methods are used as a surrogate for the total uncertainty involved in the flood frequency estimation when no discharge data is available locally. The results indicate that flood predictions in ungauged catchments are, on average, less accurate in arid than in humid climates and more accurate in large than in small catchments. There is also a tendency towards a lower performance of regressions as compared to other methods when they are applied to the same region, while geostatistical methods generally tend to perform better than other methods. For the particular case of arid catchments, index methods yield signi cantly lower performances than regression methods or geostatistical approaches. When flood records are present, design values are usually estimated by ood frequency analysis, i.e., by applying the statistical theory of extreme values to obtain a probability distribution function that describes the flood regime for a certain location. If the flood frequency estimation for an entire region is considered, the choice of a common statistical model for the entire area, also called parent distribution, becomes the rst step in approaches such as the index ood method. Chapter 3 deals with the uncertainty associated with the flood frequency model choice, by addressing the question of the existence of a parent flood frequency distribution at a European scale. A simple exploratory analysis of a newly compiled database of L-moment ratios of ood annual maximum series from 4105 catchments suggests the suitability of the Generalised Extreme Value (GEV) distribution as a pan-European ood frequency distribution. However, more detailed Monte Carlo simulations show that the GEV model underestimates the variability in terms of sample skewness and kurtosis present in the data, and particularly fails to represent the kurtosis dispersion for longer sample sizes and medium to high skewness values. Therefore, the GEV distribution was rejected in a statistical hypothesis testing framework as a single pan-European parent distribution for annual flood maxima. The results presented in this chapter indicate that one single statistical model may not be able to t the entire variety of ood processes present at a European scale. Chapter 4 further investigates the catchment and climatic factors controlling European flood regimes and their eects on the underlying flood frequency distributions. In particular, the uncertainty in statistical model choice is linked to catchment size and mean annual precipitation (MAP) using flood data from a total of 813 catchments with more than 25 years of record from Austria, Italy and Slovakia. Results shows that the GEV distribution provides a better representation for regionally averaged values of sample L-moment ratios than the other distributions considered, for catchments with medium to high MAP independently of catchment area, while the three-parameter lognormal distribution is a more appropriate choice for the drier (lower MAP) intermediate-sized catchments, which exhibit higher skewnesses. The results presented in this chapter could be seen as a rst attempt at de ning a set of "process-driven" regional parent ood frequency distributions in a European context. After analysing the uncertainties in ood frequency across dierent spatial scales in Chapters 2 to 4, Chapter 5 deals with the flood frequency estimation at a particular location in space, but extends the temporal scale of the uncertainties several centuries into the past by introducing historical flood records in the analysis. Knowledge about the historical flood regime is useful because it gives additional information that may improve the estimates of extreme discharges with high return periods and, additionally, may reduce the uncertainty in the estimates. In most practical cases, the information related to historical floods is given in a non-precise manner. This chapter presents a new approach to dealing with the imprecision present in historical floods, that links the descriptions in historical records to fuzzy numbers representing discharges. These fuzzy historical discharges are then introduced in a formal Bayesian inference framework to obtain a fuzzy version of the flood frequency curve, by combining the fuzzy historical flood events and the instrumental data for a given location. Two case studies are selected from the historical literature, representing dierent facets of the fuzziness typically present in the historical sources. The results are given in the form of the fuzzy estimates of the flood frequency curves together with the fuzzy credibility bounds for these curves. The presented fuzzy Bayesian inference framework provides a exible methodology to propagate, in an explicit, way the imprecision from the historical records to the flood frequency estimate, which allows assessing the eects of incorporating non-precise historical information in the flood frequency regime estimation. The endings presented in this thesis help better characterize and quantify dierent facets of uncertainty involved in the flood frequency estimation process. While these dierent facets of uncertainty are usually lumped together, the present work aims at throwing light at possible sources of these uncertainties, by analysing the model and data related aspects that constrain them, and by de ning the characteristic spatial and temporal scales under which they operate. The results of this thesis have implications for both hydrological understanding, and applied engineering hydrology. On the one hand, linking uncertainties in flood frequency estimation with hydrological and climatological indicators helps identify regions where an improved hydrological process understanding in needed. On the other hand, an improved quanti cation of the uncertainties helps in obtaining more robust and reliable design decisions and flood risk zones.