Technical Summary: A Survey on SAR Ship Classification using Deep Learning

1. Problem Statement

Maritime surveillance is critical for global trade, security, and environmental protection. While Automatic Identification System (AIS) data provides vessel details, it is not always available or reliable. Synthetic Aperture Radar (SAR) offers a complementary solution, capable of imaging ships in all weather and lighting conditions. However, ship classification from SAR imagery presents unique challenges:

• Data Limitations: SAR images often have limited spatial resolution (e.g., Sentinel-1 at ~20m), rendering vessels as few pixels. Furthermore, high-quality labeled datasets are scarce.
• Class Imbalance: Datasets like OpenSARShip and FUSAR-Ship exhibit severe long-tail distributions, where categories like "Cargo" dominate (up to 61–81% of data), while minority classes (e.g., fishing vessels) are underrepresented.
• Sensor Characteristics: SAR images are single-channel (unlike optical RGB), suffer from speckle noise, and contain artifacts like sidelobes. They also possess unique properties such as polarization (VV, VH, HH, HV) and phase information that are often underutilized.
• Model Generalization: Deep Learning (DL) models designed for optical imagery often fail to generalize to SAR data without specific architectural or training adaptations.

While existing surveys cover SAR ship detection, a dedicated comprehensive review of SAR ship classification using DL was absent. This paper addresses that gap.

2. Methodology

The authors conducted a systematic literature review following a rigorous protocol:

• Search Strategy: A query string combining "Deep Learning/Neural Network," "Ship/Vessel Classification," and "SAR" was applied to major databases (Web of Science, IEEE Xplore, Scopus, etc.).
• Selection Process: From an initial 1,787 records, duplicates were removed, and titles/abstracts were screened. Full-text assessment of 187 papers led to the inclusion of 74 representative studies published between 2017 and 2025.
• Inclusion Criteria: Papers had to focus on ship classification (not just detection) using SAR imagery, employ DL techniques, report quantitative metrics (F1 score and/or Overall Accuracy), and be published in English after 2010.
• Taxonomy Construction: The selected studies were categorized into a four-dimensional taxonomy:
  1. DL Architectures: Ranging from classic CNNs (AlexNet, VGG) to Residual networks, Transformers, and custom lightweight designs.
  2. Datasets: Primarily OpenSARShip and FUSAR-Ship, alongside proprietary or satellite-specific data (TerraSAR-X, COSMO-SkyMed, Gaofen-3).
  3. Image Augmentation: Geometric (rotation, flipping), noise injection, and generative approaches (GANs).
  4. Learning Techniques: Transfer learning, metric learning, feature fusion, and handling class imbalance.

3. Key Contributions

The paper provides the following specific contributions to the field:

• Systematic Review: An in-depth analysis of 74 key studies selected from 187 candidates, detailing their methodologies and performance.
• Comprehensive Taxonomy: A structured framework (Figure 2) organizing SAR ship classification research into architectures, datasets, augmentations, and learning techniques.
• Standardized Performance Reporting: A unified table reporting F1 scores and Overall Accuracy (OA) across studies, highlighting benchmarks like OpenSARShip and FUSAR-Ship while noting the limitations of cross-study comparability due to varying dataset splits.
• Empirical Trend Synthesis: Identification of effective strategies, including:
  • The efficacy of shallow CNNs under data scarcity.
  • The benefits of combining handcrafted features (e.g., HOG, geometric descriptors) with deep features.
  • The importance of polarization fusion (VV/VH) and multi-scale processing.
  • The utility of transfer learning and domain adaptation (e.g., Optical-to-SAR).
• Design Guidelines: Practical recommendations for model architecture, hyperparameter tuning (epochs, batch size, learning rate), and handling class imbalance.
• Future Directions: A roadmap for addressing open issues, including the need for SAR-specific architectures, improved interpretability, and standardized metrics.

4. Key Results and Findings

The survey analyzes the evolution and performance of DL techniques in this domain:

• Architectures: While deep models like ResNet and VGG are widely used, lightweight and custom architectures (e.g., LN-SCNet, SBNN) are gaining traction for efficiency. Hybrid models (CNN-LSTM) and attention mechanisms (SE, CBAM) show promise in capturing sequential dependencies and focusing on relevant features.
• Handcrafted Features: Integrating traditional features (HOG, LBP, geometric dimensions from AIS) with DL features consistently improves performance, particularly in data-scarce scenarios. This fusion helps models learn robust representations when pure DL approaches struggle with noise or limited samples.
• SAR Properties: Exploiting SAR-specific attributes is crucial.
  • Polarization: Dual-polarization (VV/VH) fusion generally outperforms single-polarization approaches.
  • Complex Data: Recent works utilize complex-valued CNNs to leverage phase information, improving classification of moving targets.
  • Scattering: Incorporating scattering center features or suppressing sidelobes enhances model robustness.
• Training Strategies:
  • Transfer Learning: Fine-tuning pre-trained models (from ImageNet or optical datasets) on SAR data is a standard and effective practice, especially for small datasets.
  • Data Augmentation: Techniques like rotation, flipping, and GAN-based generation are essential to mitigate data scarcity.
  • Imbalance Handling: Strategies include tailored loss functions (e.g., Focal Loss, Seesaw Loss), curriculum learning, and specific train-test splitting ratios based on minority classes.
• Performance: Reported accuracies vary significantly based on the number of classes and dataset quality. High accuracies (>90%) are often reported on 3-class subsets of OpenSARShip or FUSAR-Ship, while performance drops considerably as the number of classes increases or when dealing with highly imbalanced, fine-grained categories.

5. Significance and Claims

The paper positions itself as a foundational resource for researchers and practitioners in maritime surveillance. Its significance lies in:

• Bridging a Gap: It is the first dedicated survey to systematically review SAR ship classification (distinct from detection), providing a clear overview of the state-of-the-art.
• Practical Utility: By synthesizing empirical trends and offering design considerations (e.g., hyperparameter ranges, architecture choices), it serves as a guide for developing robust classification systems.
• Highlighting Challenges: It explicitly identifies critical bottlenecks—data scarcity, class imbalance, lack of standardized metrics, and the "black box" nature of DL models—urging the community to address them.
• Future Outlook: The paper advocates for the development of SAR-specific DL architectures, the integration of handcrafted and deep features, the adoption of standardized reporting metrics (F1 and OA), and increased interdisciplinary collaboration to advance maritime monitoring capabilities.

The authors conclude that while DL has significantly advanced SAR ship classification, achieving greater robustness, accuracy, and transparency requires concerted efforts in data curation, model interpretability, and the creation of SAR-tailored learning frameworks.