Abstract
This thesis addresses the limitations of ship classification in single-polarisation synthetic aperture radar (SAR) imagery by developing three novel methodologies that leverage the distinctive geometric, radiometric, phase, and frequency characteristics of the data. First, a confidence-aware, feature-based ship classification framework is constructed. This is achieved through a novel set of handcrafted features, derived from ship contours. Subsequently, the use of information entropy as a metric for quantifying prediction confidence is demonstrated. The findings show that the proposed features outperform existing handcrafted features and are competitive with deep learning approaches. Additionally, the entropy is observed to enable the systematic categorisation of results into distinct confidence levels, which are highly correlated with classification accuracy. Second, the value of the traditionally discarded phase component is investigated. A novel, lightweight convolutional neural network (CNN) is designed to effectively process phase data. To complement this, a comprehensive interpretability analysis using Shapley additive explanations (SHAP) is performed to quantitatively characterise the phase’s contribution. It is demonstrated that specialised integration of the phase results in a significant improvement to classification performance. Furthermore, the phase is shown to provide a complementary source of information that is attributed a similar level of importance as the amplitude. Third, the use of sub-look decomposition (SD) to generate additional input data for ship classification is examined. A rigorous theoretical analysis of processing-induced artefacts in terrain observation with progressive scans synthetic aperture radar (TOPSAR) imagery is presented. In parallel, a novel, parameter-efficient CNN is proposed to effectively fuse the resulting heterogeneous data sources. The results establish the importance of deramping TOPSAR data prior to any convolution-based operation. Moreover, it is evidenced that lightweight, custom-designed models can achieve accuracies comparable to much larger, pre-trained networks by maximising the intrinsic information within the SAR data. Collectively, the work presented underscores the importance of adapting features and models to the fundamental properties of SAR to enhance ship classification.