Machine learning for synthetic aperture radar maritime applications

The application of novel Artificial Intelligence (AI) algorithms and techniques in Remote Sensing (RS) is leading to the attainment of unprecedented results; continuous improvement in machine learning and parallel computing architectures represents an important opportunity for research to explore new frontiers and achieve state-of-the-art performances in many Remote Sensing ap plications. The use of multiple sensors, operating at different wavelengths and acquiring with various technologies, renders the Remote Sensing field of study a perfect area in which to ap ply the latest machine learning techniques. The imaging nature of the used sensors and their Multispectral/Hyperspectral capabilities give the opportunity to take spatial, textural and spectral features into account (e.g. with the use of recent machine learning architectures such as Convolu tional/Capsule Neural Networks, to name a few), while revisit time frequency enables the study of temporal dimension-related features (e.g. with the use of Recurrent Neural Networks). Within this framework, research on the application of latest machine learning technologies to data collected by a unique active sensor such as the Synthetic Aperture Radar (SAR), able to acquire regardless of weather or solar illumination conditions and capable of centimeter to millimeter resolution, represents a challenging and relatively new topic. SAR multifrequency and polarimetric capabilities provide a large amount of information related to scattering mechanisms at different wavelengths and polarizations, enabling exclusive and refined studies, albeit implicating more considerations on the underlying data. Moreover, the large amount of daily collected data from spaceborne and airborne missions also entails the possibility to use data-fusion and dimensionality reduction techniques, e.g. through manifold learning approaches. Based on these premises, the scientific community is called to tackle new challenges in the application of machine learning to RS and to SAR data in particular. Main contribution of this thesis consists in the study of the application of latest deep learning techniques to SAR maritime applications, such as automatic shoreline extraction, ships and oil slicks classification; obtained results from research studies presented in this thesis are to be regarded as relevant, establishing state-of-the-art findings under certain circumstances.