Novel Architectures for Radar Sounder Signals Segmentation: From Convolutional Neural Networks to Quantum-Enhanced Networks

Semantic segmentation has emerged as a cornerstone in the analysis of Earth observation and Planetary exploration data for enabling detailed interpretation of complex datasets. This thesis focuses on advancing the state of the art in semantic segmentation methods with a particular emphasis on subsurface information extraction from radar sounder (RS) data. RSs are essential tools in subsurface explorations that present unique challenges due to their complex signal characteristics. They transmit linearly modulated electromagnetic (EM) pulses in the nadir directions to profile the subsurface targets by coherently integrating the backscattered echoes received by the RSs antenna. The research encapsulated in this thesis introduces a series of methodologically diverse and technically robust frameworks designed to elevate the performance and adaptability of semantic segmentation models which can be adopted for upcoming planetary exploration missions such as JUICE, EnVision, etc. The proposed innovative methodologies are grounded in hybrid supervised deep learning architectures, unsupervised feature learning frameworks, and harnessing quantum machine learning frameworks to automatically segment geological units in the cryosphere subsurface. Firstly, we developed a supervised deep learning architecture by synergistically integrating the strength of convolutional neural networks (CNNs) in capturing local spatial contexts and transformers in capturing global spatial contexts embedded in radargrams for efficiently segmenting the radargrams. Since it is difficult to obtain accurate labelled information in cryospheric regions due to miscellaneous logistical constraints, we constructed an unsupervised feature learning framework to facilitate subsurface feature extraction without incorporating any labelled information during training. The proposed unsupervised method incorporated self-supervised Vision Transformers (ViTs) coupled with contrastive feature learning to overcome the scarcity of labelled data. Thirdly, we conducted the first-ever exploration of emerging quantum machine learning frameworks for RS signal segmentation. Particularly, the thesis ventured into integrating quantum computing paradigms with classical architectures by crafting hybrid quantum-classical CNNs to explore the potential of quantum feature maps in radargram segmentation. Comprehensive evaluations on diverse radargram datasets underline the efficacy of the proposed methodologies. The results reveal not only enhanced segmentation accuracy but also improved resilience to noise and variations in radar data, offering a transformative leap from conventional approaches. Beyond the immediate context of RS data, the research reflects on its broader applicability to other domains such as Synthetic Aperture Radar (SAR). Through these contributions, this thesis sets a foundation for future research in classical and quantum paradigms as well as their combination for complex Earth observation and Planetary exploration challenges. It aspires to contribute not only the larger scientific endeavour of understanding and decoding the complex data landscapes of Earth observation and Planetary explorations but also to the field of semantic segmentation. Finally, this work seeks to inspire a new wave of research that bridges technological frontiers, paving the way for more efficient, versatile, and insightful models in the years to come.