Machine Learning and Pattern Recognition Methods for Remote Sensing Image Registration and Fusion

In the last decade, the remote sensing world has dramatically evolved. New types of sensor, each one collecting data with possibly different modalities, have been designed, developed, and deployed. Moreover, new missions have been planned and launched, aimed not only at collecting data of the Earth's surface, but also at acquiring planetary data in support of the study of the whole Solar system. Indeed, such a variety of technologies highlights the need for automatic methods able to effectively exploit all the available information. In the last years, lot of effort has been put in the design and development of advanced data fusion methods able to extract and make use of all the information available from as many complementary information sources as possible. Indeed, the goal of this thesis is to present novel machine learning and pattern recognition methodologies designed to support the exploitation of diverse sources of information, such as multisensor, multimodal, or multiresolution imagery. In this context, image registration plays a major role as is allows bringing two or more digital images into precise alignment for analysis and comparison. Here, image registration is tackled using both feature-based and area-based strategies. In the former case, the features of interest are extracted using a stochastic geometry model based on marked point processes, while, in the latter case, information theoretic functionals and the domain adaptation capabilities of generative adversarial networks are exploited. In addition, multisensor image registration is also applied in a large scale scenario by introducing a tiling-based strategy aimed at minimizing the computational burden, which is usually heavy in the multisensor case due to the need for information theoretic similarity measures. Moreover, automatic change detection with multiresolution and multimodality imagery is addressed via a novel Markovian framework based on a linear mixture model and on an ad-hoc multimodal energy function minimized using graph cuts or belied propagation methods. The statistics of the data at the various spatial scales is modelled through appropriate generalized Gaussian distributions and by iteratively estimating a set of virtual images, at the finest resolution, representing the data that would have been collected in case all the sensors worked at that resolution. All such methodologies have been experimentally evaluated with respect to different datasets, and with particular focus on the trade-off between the achievable performances and the demands in terms of computational resources. Moreover, such methods are also compared with state-of-the-art solutions, and are analyzed in terms of future developments, giving insights to possible future lines of research in this field.