Deep Learning Methods for SAR Ship Classification under Data Scarcity, Class Imbalance, and Limited Resolution

Deep learning has become a dominant approach in processing remote sensing data, included Synthetic Aperture Radar imagery (SAR) for ship classification, yet operational deployments remain constrained by three practical factors: limited labelled datasets, long-tailed class distributions, and altered information content in SAR imaging due to noise and limited spatial resolution. The objective of this thesis is the systematic analysis of these issues, and the development of deep learning-based countermeasures aiming at the improvement of the ship classification task through the processing of SAR images, under real experimental conditions. Based on a thorough assessment of the related scientific literature, the thesis first goes through the background and taxonomy for SAR ship classification, covering the currently available datasets, the mostly exploited core architectures (convolutional and attention-based), and the many challenges potentially emerging during the model preparation phases. Building on this foundation, the analysis focuses on data scarcity, dataset quality, and domain shift. Inter-dataset generalisation has been assessed under a shared three-class setting, showing that cross-dataset performance can degrade substantially, even when label sets overlap. In addition, the sensitivity of deep learning algorithms to training objectives across datasets has been investigated, showing that the relative behaviour of loss functions can change under dataset shift. This reinforces the idea that the objective function is not a fixed default, but must be considered a critical component of the model, to be carefully adapted and optimized. This work also addresses class imbalance and training dynamics in the case of long-tailed SAR ship datasets, demonstrating that overall accuracy can remain deceptively stable while minority-class performance and stability remain poor. To counteract this, the thesis develops imbalance-aware training strategies based on curriculum learning and weighted losses, and introduces an imbalance-aware evaluation score designed to complement overall accuracy and macro-F1 in long-tailed regimes. Further analysis concerns feature-space oversampling as a representation-level alternative to naive image duplication. It is shown that enriching minority regions of the embedding space can improve minority-class recognition. Afterwards, data quality issues are considered, by studying super-resolution (SR) for SAR ship classification. First, the relationship between image fidelity improvements and downstream recognition is analysed, demonstrating that gains in common fidelity metrics do not reliably translate into improved classification performance. Motivated by this mismatch, a second approach is presented, proposing a classification-aware SR framework that aligns SR optimisation with recognition objectives through stage-wise training and loss design, supported by extensive quantitative comparisons and qualitative analyses. Overall, this thesis provides a unified view of robustness for SAR ship classification along three coupled axes: scarcity and dataset shift, imbalance and training dynamics, and pixel-level information content. The results support practical guidelines for evaluation and model design, including the use of class-sensitive metrics, cross-dataset testing when claiming robustness, imbalance-aware training strategies when long-tailed distributions dominate, and task-level validation for preprocessing interventions such as super-resolution. Together, the contributions aim to support the development of SAR ship classifiers that remain reliable when training data are scarce, heterogeneous, and imbalanced, and when the input pixel contain limited discriminative detail.