In the last decades, the increasing number of new generation satellite images characterized by a better spectral, spatial and temporal resolution with respect to the past has provided unprecedented source of information for monitoring climate changes.To exploit this wealth of data, powerful and automatic methods to analyze remote sensing images need to be implemented. Accordingly, the objective of this thesis is to develop advanced methods for the analysis of multitemporal multispectral remote sensing images to support climate change applications. The thesis is divided into two main parts and provides four novel contributions to the state-of-the-art. In the first part of the thesis, we exploit multitemporal and multispectral remote sensing data for accurately monitoring two essential climate variables. The first contribution presents a method to improve the estimation of the glacier mass balance provided by physically-based models. Unlike most of the literature approaches, this method integrates together physically-based models, remote sensing data and in-situ measurements to achieve an accurate and comprehensive glacier mass balance estimation. The second contribution addresses the land cover mapping for monitoring climate change at high spatial resolution. Within this work, we developed two processing chains: one for the production of a recent (2019) static high resolution (10 m) land cover map at subcontinental scale, and the other for the production of a long-term record of regional high resolution (30 m) land cover maps. The second part of this thesis addresses the common challenges faced while performing the analysis of multitemporal multispectral remote sensing data. In this context, the third contribution deals with the multispectral images cloud occlusions problem. Differently from the literature, instead of performing computationally expensive cloud restoration techniques, we study the robustness of deep learning architectures such as Long Short Term Memory classifier to cloud cover. Finally, we address the problem of the large scale training set definition for multispectral data classification. To this aim, we propose an approach that leverages on available low resolution land cover maps and domain adaptation techniques to provide representative training sets at large scale. The proposed methods have been tested on Sentinel-2 and Landsat 5, 7, 8 multispectral images. Qualitative and quantitative experimental results confirm the effectiveness of the methods proposed in this thesis.

Ce travail est motivé par la spectroscopie de photoélectrons et l'étude de la cinématique des galaxies où les données correspondent respectivement à une séquence temporelle de spectres et à une image multispectrale. L'objectif est d'estimer les caractéristiques (amplitude, position spectrale et paramètre de forme) des raies présentes dans les spectres, ainsi que leur évolution au sein des données. Dans les applications considérées, cette évolution est lente puisque deux spectres voisins sont souvent très similaires : c'est une connaissance a priori qui sera prise en compte dans les méthodes développées. Ce problème inverse est abordé sous l'angle de la séparation de sources retardées, où les spectres et les raies sont attribués respectivement aux mélanges et aux sources. Les méthodes de l'état de l'art sont inadéquates car elles supposent la décorrélation ou l'indépendance des sources, ce qui n'est pas le cas. Nous tirons parti de la connaissance des sources pour les modéliser par une fonction paramétrique. Nous proposons une première méthode de moindres carrés alternés : les paramètres de formes sont estimés avec l'algorithme de Levenberg-Marquardt, tandis que les amplitudes et les positions sont estimées avec un algorithme inspiré d'Orthogonal Matching Pursuit. Une deuxième méthode introduit un terme de régularisation pour prendre en compte l'évolution lente des positions; un nouvel algorithme d'approximation parcimonieuse conjointe est alors proposée. Enfin, une troisième méthode contraint l'évolution des amplitudes, positions et paramètres de forme par des fonctions B-splines afin de garantir une évolution lente conforme au physique des phénomènes observés. Les points de contrôle des B-splines sont estimés par un algorithme de moindre carrés non-linéaires. Les résultats sur des données synthétiques et réelles montrent que les méthodes proposées sont plus efficaces que les méthodes de l'état de l'art et aussi efficaces qu'une méthode bayésienne adaptée au problème mais avec un temps de calcul sensiblement réduit. / This work is motivated by photoelectron spectroscopy and the study of galaxy kinematics where data respectively correspond to a temporal sequence of spectra and a multispectral image. The objective is to estimate the characteristics (amplitude, spectral position and shape) of peaks embedded in the spectra, but also their evolution within the data. In the considered applications, this evolution is slow since two neighbor spectra are often very similar: this a priori knowledge that will be taken into account in the developed methods. This inverse problem is approached as a delayed source separation problem where spectra and peaks are respectively associated with mixtures and sources. The state-of-the-art methods are inadequate because they suppose the source decorrelation and independence, which is not the case. We take advantage of the source knowledge in order to model them by a parameterized function. We first propose an alternating least squares method: the shape parameters are estimated with the Levenberg-Marquardt algorithm, whilst the amplitudes and positions are estimated with an algorithm inspired from Orthogonal Matching Pursuit. A second method introduces a regularization term to consider the delay slow evolution; a new joint sparse approximation algorithm is thus proposed. Finally, a third method constrains the evolution of the amplitudes, positions and shape parameters by B-spline functions to guarantee their slow evolution. The B-spline control points are estimated with a non-linear least squares algorithm. The results on synthetic and real data show that the proposed methods are more effective than state-of-the-art methods and as effective as a Bayesian method which is adapted to the problem. Moreover, the proposed methods are significantly faster.