The development of flexible and interpretable statistical methods is necessary in order to provide appropriate risk assessment measures for extreme events and natural disasters. In this thesis, we address this challenge by contributing to the developing research field of Extreme-Value Theory. We initially study the performance of existing parametric and non-parametric estimators of extremal dependence for multivariate maxima. As the dimensionality increases, non-parametric estimators are more flexible than parametric methods but present some loss in efficiency that we quantify under various scenarios. We introduce a statistical tool which imposes the required shape constraints on non-parametric estimators in high dimensions, significantly improving their performance. Furthermore, by embedding the tree-based max-stable nested logistic distribution in the Bayesian framework, we develop a statistical algorithm that identifies the most likely tree structures representing the data's extremal dependence using the reversible jump Monte Carlo Markov Chain method. A mixture of these trees is then used for uncertainty assessment in prediction through Bayesian model averaging. The computational complexity of full likelihood inference is significantly decreased by deriving a recursive formula for the nested logistic model likelihood. The algorithm performance is verified through simulation experiments which also compare different likelihood procedures. Finally, we extend the nested logistic representation to the spatial framework in order to jointly model multivariate variables collected across a spatial region. This situation emerges often in environmental applications but is not often considered in the current literature. Simulation experiments show that the new class of multivariate max-stable processes is able to detect both the cross and inner spatial dependence of a number of extreme variables at a relatively low computational cost, thanks to its Bayesian hierarchical representation. These innovative methods and models are implemented to study the concentration maxima of various air pollutants and how these are related to extreme weather conditions for a number of sites in California, one of the most populated and polluted states of the US. As a result, we provide comprehensive measures of air quality that can be used by communities and policymakers worldwide to better assess and manage the health, environmental and financial impacts of air pollution extremes.

The statistical modeling of extreme natural hazards is becoming increasingly important due to climate change, whose effects have been increasingly visible throughout the last decades. It is thus crucial to understand the dependence structure of rare, high-impact events over space and time for realistic risk assessment. For spatial extremes, max-stable processes have played a central role in modeling block maxima. However, the spatial tail dependence strength is persistent across quantile levels in those models, which is often not realistic in practice. This lack of flexibility implies that max-stable processes cannot capture weakening dependence at increasingly extreme levels, resulting in a drastic overestimation of joint tail risk. To address this, we develop new dependence models in this thesis from the class of max-infinitely divisible (max-id) processes, which contain max-stable processes as a subclass and are flexible enough to capture different types of dependence structures. Furthermore, exact simulation algorithms for general max-id processes are typically not straightforward due to their complex formulations. Both simulation and inference can be computationally prohibitive in high dimensions. Fast and exact simulation algorithms to simulate max-id processes are provided, together with methods to implement our models in high dimensions based on the Vecchia approximation method. These proposed methodologies are illustrated through various environmental datasets, including air temperature data in South-Eastern Europe in an attempt to assess the effect of climate change on heatwave hazards, and sea surface temperature data for the entire Red Sea. In another application focused on assessing how the spatial extent of extreme precipitation has changed over time, we develop new time-varying $r$-Pareto processes, which are the counterparts of max-stable processes for high threshold exceedances.

Les extrêmes neigeux sont parmi les risques naturels les plus dangereux dans les régions montagneuses. Les processus max-stables, qui relient statistique des valeurs extrêmes et géostatistique, offrent un cadre approprié pour les étudier. Deux questions importantes concernant la dépendance spatiale des extrêmes sont traitées dans cette thèse à travers les cas des chutes et des hauteurs de neige dans les Alpes françaises : la sélection de modèle et la non-stationnarité temporelle. Nous utilisons pour cela deux jeux de données de maxima hivernaux de chutes de neige (90 stations de 1958 à 2013) et de hauteurs de neige (82 stations de 1970 à 2013). Nous décrivons d'abord une procédure de validation-croisée appropriée pour évaluer les capacités des processus max-stables à capturer la structure de dépendance des extrêmes spatiaux. Nous mettons en exergue trois processus max-stables pour leur aptitude à modéliser la dépendance spatiale des chutes de neige extrêmes : les processus de Brown-Resnick, géométrique gaussien et extrémal-t. Les performances de ces trois modèles sont extrêmement similaires, quel que soit le nombre de stations ou d'années. Ensuite, nous présentons une approche par fenêtre glissante pour évaluer l'évolution temporelle de la dépendance des extrêmes spatiaux. Nous montrons ainsi que les chutes de neige extrêmes ont tendance à être de moins en moins dépendantes spatialement. Nous montrons que cela est dû à une augmentation de la température provoquant une baisse du ratio neige/pluie. Il existe aussi un effet d'intensité avec des extrêmes moins dépendants à cause d'une baisse du cumul hivernal de chutes de neige. Enfin, nous présentons la première utilisation de processus max-stables avec des tendances temporelles dans la structure de dépendance spatiale. Cette approche est appliquée aux maxima de hauteurs de neige modélisés par un processus de Brown-Resnick. Nous montrons que leur dépendance spatiale est impactée par le changement climatique d'une manière similaire que celle des chutes de neige extrêmes. / Extreme snowfall and extreme snow depths are among the most dangerous hazards in the mountainous regions. Max-stable processes, which connect extreme value statistics and geostatistics by modeling the spatial dependence of extremes, offer a suitable framework to deal with. Two challenging issues concerning spatial dependence of extremes are broached in this thesis through the examples of snowfall and snow depths in the French Alps: model selection and temporal nonstationarity. We process two winter maxima data sets of 3-day snowfall (90 stations from 1958 to 2013) and snow depths (82 stations from 1970 to 2013). First, we introduce a leave-two-out cross-validation procedure appropriate for evaluating the predictive ability of max-stable processes to model the dependence structure of spatial extremes. We compare five of the most commonly used max-stable processes, using as a case study the snowfall maxima data set. This approach allows us to show that the extremal-t, geometric Gaussian and Brown-Resnick processes are able to represent as well the structure of dependence of the data, regardless of the number of stations or years. Then, we show, using a data-based approach allowing to make minimal modeling assumptions, that snowfall extremes tended to become less spatially dependent over time, with the dependence range reduced roughly by half during the study period. We demonstrate that this is attributable at first to the increase in temperature and its major control on the snow/rain partitioning. A magnitude effect, with less dependent extremes due to a decrease in winter cumulated snowfall, also exists. Finally, we tackle the first-ever use of max-stable processes with temporal trends in the spatial dependence structure. This approach is applied to snow depth winter maxima modeled by a Brown-Resnick process. We show that the spatial dependence of extreme snow depths is impacted by climate change in a similar way to that has been observed for extreme snowfall.

La prédiction de futurs évènements extrêmes est d’un grand intérêt dans de nombreux domaines tels que l’environnement ou la gestion des risques. Alors que la théorie des valeurs extrêmes univariées est bien connue, la complexité s’accroît lorsque l’on s’intéresse au comportement joint d’extrêmes de plusieurs variables. Un intérêt particulier est porté aux évènements de nature spatiale, définissant le cadre d’un nombre infini de dimensions. Sous l’hypothèse que ces évènements soient marginalement extrêmes, nous focalisons sur la structure de dépendance qui les lie. Dans un premier temps, nous faisons une revue des modèles paramétriques de dépendance dans le cadre multivarié et présentons différentes méthodes d’estimation. Les processus maxstables permettent l’extension au contexte spatial. Nous dérivons la loi en dimension finie du célèbre modèle de Brown- Resnick, permettant de faire de l’inférence par des méthodes de vraisemblance ou de vraisemblance composée. Nous utilisons ensuite des lois asymétriques afin de définir la représentation spectrale d’un modèle plus large : le modèle Extremal Skew-t, généralisant la plupart des modèles présents dans la littérature. Ce modèle a l’agréable propriété d’être asymétrique et non-stationnaire, deux notions présentées par les évènements environnementaux spatiaux. Ce dernier permet un large spectre de structures de dépendance. Les indicateurs de dépendance sont obtenus en utilisant la loi en dimension finie.Enfin, nous présentons une méthode d’estimation non-paramétrique par noyau pour les queues de distributions et l’appliquons à la sélection de modèles. Nous illustrons notre méthode à partir de l’exemple de modèles climatiques. / Projection of future extreme events is a major issue in a large number of areas including the environment and risk management. Although univariate extreme value theory is well understood, there is an increase in complexity when trying to understand the joint extreme behavior between two or more variables. Particular interest is given to events that are spatial by nature and which define the context of infinite dimensions. Under the assumption that events correspond marginally to univariate extremes, the main focus is then on the dependence structure that links them. First, we provide a review of parametric dependence models in the multivariate framework and illustrate different estimation strategies. The spatial extension of multivariate extremes is introduced through max-stable processes. We derive the finite-dimensional distribution of the widely used Brown-Resnick model which permits inference via full and composite likelihood methods. We then use Skew-symmetric distributions to develop a spectral representation of a wider max-stable model: the extremal Skew-t model from which most models available in the literature can be recovered. This model has the nice advantages of exhibiting skewness and nonstationarity, two properties often held by environmental spatial events. The latter enables a larger spectrum of dependence structures. Indicators of extremal dependence can be calculated using its finite-dimensional distribution. Finally, we introduce a kernel based non-parametric estimation procedure for univariate and multivariate tail density and apply it for model selection. Our method is illustrated by the example of selection of physical climate models.

No description available.

No description available.