Invatarea profunda pentru date SAR in prezenta esantintelor adversiale / Deep learning for SAR data in presence of adversarial samples

SAR sensors, due to their unique capabilities, play an essential role in ensuring worldwide and uninterrupted monitoring. However, SAR data have a high degree of complexity; they are Complex-Valued (CV) multidimensional signals with particular properties and inherent adversarial effect induced by the coherent imaging mode and the observed scene scattering process. In this thesis we aim to provide deep learning-based solutions for SAR data in pretense of adversarial samples, considering their unique characteristics. We focus on CV networks for this purpose to fully exploit the amplitude and phase components of SAR data. A critical analysis is carried out on generative models and their ability to model complex SAR data distribution in presence of adversarial samples. Later, big EO data mining of the semantic content is investigated to generate a CV semantically annotated dataset, called S1SLC_CVDL, for training CV deep networks. Moving forward, the superiority of the CV deep architectures for various SAR applications, including classification, is demonstrated. Our findings show that the CV networks can achieve higher classification accuracies with fewer training samples and shallower architectures (less computational cost); and preserve the complex coherence, physical data model, Doppler attributes, and original properties of SAR data. Considering their enormous potential, we delve into the more practical applications by employing CV networks to engineer a novel data compression approach, utilizing CV autoencoders, tailored for compressing raw SAR data. The demonstrated capabilities of the CV deep architectures in this study pave the way for the future development of physics-aware trustworthy explainable artificial intelligent methods for SAR data.