Improving SAR Wake Detection Performance with Granular Training data: Classification of Kelvin wake and Turbulent wake

As the importance of maritime domain awareness (MDA) increases, the detection of ships utilizing satellite imagery has shown significant advances. In particular, Synthetic Aperture Radar (SAR) imagery can be used for ship detection regardless of weather conditions or time of day. However, the detection of small or non-metallic vessels can be challenging due to their low radar cross-section and the presence of sea clutter. To overcome these limitations, wake detection, which takes advantage of the extensive and long-lasting patterns generated by vessels, has received significant attention. Recent studies have actively explored the application of AI techniques to wake detection based on SAR imagery to enhance detection performance. However, the effectiveness of AI-based models is highly dependent on the quality of the training data, and publicly available datasets classify all wakes under a single category, thereby not fully capturing the diverse characteristics of wakes. To address this, this study introduces a granular labeling approach that distinguishes between the Kelvin wake and turbulent wake, aiming to improve the performance of SAR wake detection. Dataset analysis revealed that turbulent wakes are predominant and more easily observed in SAR imagery, while kelvin wakes are relatively rare and difficult to detect. Consequently, a separate class for Turbulent wake was created, which led to improved model performance. To resolve the class imbalance issue associated with the Kelvin wake, the simulated data was integrated into the dataset. This allowed the model to learn the unique features of both wake types more effectively. The combination of granular classification and simulated data resulted in a modest improvement in the overall accuracy of wake detection. The findings suggest that granular wake classification and data augmentation play an important role in improving the performance of SAR-based wake detection.