In the era of unprecedented climatic, geomorphologic, environmental, and anthropogenic changes in the Earth, a global-scale, long-term, and continues monitoring via Earth Observation (EO) sensors is imperative. Among EO sensors, Synthetic Aperture Radar (SAR) systems stand out due to their day and night observation capability and immunity to atmospheric conditions, and play an essential role in ensuring uninterrupted worldwide monitoring. However, SAR data have a high degree of complexity; they are Complex-Valued (CV) multidimensional signals with particular properties induced by the coherent imaging mode and the observed scene scattering process and the inherent adversarial effect.
Deep learning has emerged as a remarkably potent and widely adopted technique across diverse fields, showcasing its unparalleled effectiveness in tackling complex challenges, including remote sensing. This thesis is dedicated to explore novel deep learning-based solutions for SAR applications, considering unique characteristics of SAR data and capability of deep networks to learn and model data distribution of SAR data, to unveil new perspectives in this field. We delve into the CV deep architectures for this purpose to fully exploit the amplitude and phase components of SAR data. The research presented in this thesis can be classified into three parts:
In the first part, we investigate the Bayesian generative model, Latent Dirichlet Allocation, for big EO data mining of the semantic content and generate a CV semantically annotated dataset from Sentinel-1 (S1) Single Look Complex (SLC) StripMap (SM) mode products, called S1SLC_CVDL, for training CV deep networks.
Moving forward, the second part of the thesis is dedicated to the implementation of the CV networks and comprehensive analyses of these models, with respect to the particular characteristics of CV-SAR data. In this part, a wide range of operators, layers, and functions are converted into the complex domain for CV network’s implementation. Later, extensive investigations are carried out on CV deep architectures for various SAR applications, illuminating the supremacy of CV models for semantic land cover classification, data distribution modelling, complex coherence preservation, and physical attributes interpretation and retrieval from SAR data.
Acknowledging their enormous potential, in the last part, we venture into the practical and more complicated applications of the CV networks. We employ CV networks to engineer a novel data compression approach utilizing CV autoencoders, tailored for compressing SAR raw data.
The demonstrated capabilities of the CV deep architectures in this thesis unravel new perspectives in the field of CV deep architectures for SAR applications and pave the way for the future development of physics-aware CV deep networks with data distribution modelling capability for various remote sensing applications.