| Reliable mine detection in side-scan sonar (SSS) imagery can be hindered by limited datasets. This study investigates the use of Generative Adversarial Networks (GANs) to augment scarce sonar data for improving deep learning–based mine classification. Two generative models, StyleGAN3 and FastGAN, were trained on a real SSS dataset provided by the Finnish Navy to produce synthetic samples for four-class classification tasks. The generated datasets were combined with real images at varying ratios (0–100% of synthetic data used) and evaluated across four architectures: Autoencoder, Vision Transformer (ViT), Distilled ViT, and a CNN–Transformer Hybrid. Results show that StyleGAN3 augmentation, improved average classification performance by up to 2–4 percentage points, particularly in the 40–80% synthetic ratio range, while FastGAN data yielded inconsistent or negative effects at higher ratios. The Autoencoder achieved the highest overall average accuracy of 0.64, though its performance fluctuated considerably with different amounts of synthetic data augmentation. These findings indicate that in right conditions synthetic augmentation can enhance sonar-based mine detection, whereas lower-quality synthetic samples may degrade model reliability. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.