Le démélange spectral est l’un des problèmes centraux pour l’exploitation des images hyperspectrales. En raison de la faible résolution spatiale des imageurs hyperspectraux en télédetection, la surface représentée par un pixel peut contenir plusieurs matériaux. Dans ce contexte, le démélange consiste à estimer les spectres purs (les end members) ainsi que leurs fractions (les abondances) pour chaque pixel de l’image. Le but de cette thèse estde proposer de nouveaux algorithmes de démélange qui visent à améliorer l’estimation des spectres purs et des abondances. En particulier, les algorithmes de démélange proposés s’inscrivent dans le cadre du démélange non-supervisé et non-linéaire. Dans un premier temps, on propose un algorithme de démelange non-supervisé dans lequel une régularisation favorisant la parcimonie des groupes est utilisée pour identifier les spectres purs parmi les observations. Une extension de ce premier algorithme permet de prendre en compte la présence du bruit parmi les observations choisies comme étant les plus pures. Dans un second temps, les connaissances a priori des ressemblances entre les spectres à l’échelle localeet non-locale ainsi que leurs positions dans l’image sont exploitées pour construire un graphe adapté à l’image. Ce graphe est ensuite incorporé dans le problème de démélange non supervisé par le biais d’une régularisation basée sur le Laplacian du graphe. Enfin, deux algorithmes de démélange non-linéaires sont proposés dans le cas supervisé. Les modèles de mélanges non-linéaires correspondants incorporent des fonctions à valeurs vectorielles appartenant à un espace de Hilbert à noyaux reproduisants. L’intérêt de ces fonctions par rapport aux fonctions à valeurs scalaires est qu’elles permettent d’incorporer un a priori sur la ressemblance entre les différentes fonctions. En particulier, un a priori spectral, dans un premier temps, et un a priori spatial, dans un second temps, sont incorporés pour améliorer la caractérisation du mélange non-linéaire. La validation expérimentale des modèles et des algorithmes proposés sur des données synthétiques et réelles montre une amélioration des performances par rapport aux méthodes de l’état de l’art. Cette amélioration se traduit par une meilleure erreur de reconstruction des données / Spectral unmixing has been an active field of research since the earliest days of hyperspectralremote sensing. It is concerned with the case where various materials are found inthe spatial extent of a pixel, resulting in a spectrum that is a mixture of the signatures ofthose materials. Unmixing then reduces to estimating the pure spectral signatures and theircorresponding proportions in every pixel. In the hyperspectral unmixing jargon, the puresignatures are known as the endmembers and their proportions as the abundances. Thisthesis focuses on spectral unmixing of remotely sensed hyperspectral data. In particular,it is aimed at improving the accuracy of the extraction of compositional information fromhyperspectral data. This is done through the development of new unmixing techniques intwo main contexts, namely in the unsupervised and nonlinear case. In particular, we proposea new technique for blind unmixing, we incorporate spatial information in (linear and nonlinear)unmixing, and we finally propose a new nonlinear mixing model. More precisely, first,an unsupervised unmixing approach based on collaborative sparse regularization is proposedwhere the library of endmembers candidates is built from the observations themselves. Thisapproach is then extended in order to take into account the presence of noise among theendmembers candidates. Second, within the unsupervised unmixing framework, two graphbasedregularizations are used in order to incorporate prior local and nonlocal contextualinformation. Next, within a supervised nonlinear unmixing framework, a new nonlinearmixing model based on vector-valued functions in reproducing kernel Hilbert space (RKHS)is proposed. The aforementioned model allows to consider different nonlinear functions atdifferent bands, regularize the discrepancies between these functions, and account for neighboringnonlinear contributions. Finally, the vector-valued kernel framework is used in orderto promote spatial smoothness of the nonlinear part in a kernel-based nonlinear mixingmodel. Simulations on synthetic and real data show the effectiveness of all the proposedtechniques

La classe des graphes quasi-médians est une généralisation des graphes médians, ou de manière équivalente, des complexes cubiques CAT(0). L'objectif de cette thèse est d'introduire ces graphes dans le monde de la théorie géométrique des groupes. Dans un premier temps, nous étendons la notion d'hyperplan définie dans les complexes cubiques CAT(0), et nous montrons que la géométrie d'un graphe quasi-médian se réduit essentiellement à la combinatoire de ses hyperplans. Dans la deuxième partie de notre texte, qui est le cœur de la thèse, nous exploitons la structure particulière des hyperplans pour démontrer des résultats de combinaison. L'idée principale est que si un groupe agit d'une bonne manière sur un graphe quasi-médian de sorte que les stabilisateurs de cliques satisfont une certaine propriété P de courbure négative ou nulle, alors le groupe tout entier doit satisfaire P également. Les propriétés que nous considérons incluent : l'hyperbolicité (éventuellement relative), les compressions lp (équivariantes), la géométrie CAT(0) et la géométrie cubique. Finalement, la troisième et dernière partie de la thèse est consacrée à l'application des critères généraux démontrés précédemment à certaines classes de groupes particulières, incluant les produits graphés, les groupes de diagrammes introduits par Guba et Sapir, certains produits en couronne, et certains graphes de groupes. Les produits graphés constituent notre application la plus naturelle, où le lien entre le groupe et son graphe quasi-médian associé est particulièrement fort et explicite; en particulier, nous sommes capables de déterminer précisément quand un produit graphé est relativement hyperbolique. / The class of quasi-median graphs is a generalisation of median graphs, or equivalently of CAT(0) cube complexes. The purpose of this thesis is to introduce these graphs in geometric group theory. In the first part of our work, we extend the definition of hyperplanes from CAT(0) cube complexes, and we show that the geometry of a quasi-median graph essentially reduces to the combinatorics of its hyperplanes. In the second part, we exploit the specific structure of the hyperplanes to state combination results. The main idea is that if a group acts in a suitable way on a quasi-median graph so that clique-stabilisers satisfy some non-positively curved property P, then the whole group must satisfy P as well. The properties we are interested in are mainly (relative) hyperbolicity, (equivariant) lp-compressions, CAT(0)-ness and cubicality. In the third part, we apply our general criteria to several classes of groups, including graph products, Guba and Sapir's diagram products, some wreath products, and some graphs of groups. Graph products are our most natural examples, where the link between the group and its quasi-median graph is particularly strong and explicit; in particular, we are able to determine precisely when a graph product is relatively hyperbolic.

Compressed sensing has been proposed as a computationally efficient method to estimate the finite-dimensional signals. The idea is to develop an undersampling operator that can sample the large but finite-dimensional sparse signals with a rate much below the required Nyquist rate. In other words, considering the sparsity level of the signal, the compressed sensing samples the signal with a rate proportional to the amount of information hidden in the signal. In this dissertation, first, we employ compressed sensing for physical layer signal processing of directional millimeter-wave communication. Second, we go through the theoretical aspect of compressed sensing by running a comprehensive theoretical analysis of compressed sensing to address two main unsolved problems, (1) continuous-extension compressed sensing in locally convex space and (2) computing the optimum subspace and its dimension using the idea of equivalent topologies using Köthe sequence. In the first part of this thesis, we employ compressed sensing to address various problems in directional millimeter-wave communication. In particular, we are focusing on stochastic characteristics of the underlying channel to characterize, detect, estimate, and track angular parameters of doubly directional millimeter-wave communication. For this purpose, we employ compressed sensing in combination with other stochastic methods such as Correlation Matrix Distance (CMD), spectral overlap, autoregressive process, and Fuzzy entropy to (1) study the (non) stationary behavior of the channel and (2) estimate and track channel parameters. This class of applications is finite-dimensional signals. Compressed sensing demonstrates great capability in sampling finite-dimensional signals. Nevertheless, it does not show the same performance sampling the semi-infinite and infinite-dimensional signals. The second part of the thesis is more theoretical works on compressed sensing toward application. In chapter 4, we leverage the group Fourier theory and the stochastical nature of the directional communication to introduce families of the linear and quadratic family of displacement operators that track the join-distribution signals by mapping the old coordinates to the predicted new coordinates. We have shown that the continuous linear time-variant millimeter-wave channel can be represented as the product of channel Wigner distribution and doubly directional channel. We notice that the localization operators in the given model are non-associative structures. The structure of the linear and quadratic localization operator considering group and quasi-group are studied thoroughly. In the last two chapters, we propose continuous compressed sensing to address infinite-dimensional signals and apply the developed methods to a variety of applications. In chapter 5, we extend Hilbert-Schmidt integral operator to the Compressed Sensing Hilbert-Schmidt integral operator through the Kolmogorov conditional extension theorem. Two solutions for the Compressed Sensing Hilbert Schmidt integral operator have been proposed, (1) through Mercer's theorem and (2) through Green's theorem. We call the solution space the Compressed Sensing Karhunen-Loéve Expansion (CS-KLE) because of its deep relation to the conventional Karhunen-Loéve Expansion (KLE). The closed relation between CS-KLE and KLE is studied in the Hilbert space, with some additional structures inherited from the Banach space. We examine CS-KLE through a variety of finite-dimensional and infinite-dimensional compressible vector spaces. Chapter 6 proposes a theoretical framework to study the uniform convergence of a compressible vector space by formulating the compressed sensing in locally convex Hausdorff space, also known as Fréchet space. We examine the existence of an optimum subspace comprehensively and propose a method to compute the optimum subspace of both finite-dimensional and infinite-dimensional compressible topological vector spaces. To the author's best knowledge, we are the first group that proposes continuous compressed sensing that does not require any information about the local infinite-dimensional fluctuations of the signal.