Improved Data Augmentation and Cross-Modal Feature Fusion for Sonar Target Classification Under Reverberation Background

Effective sonar target classification in reverberant underwater environments remains a significant challenge due to limited sample data and complex noise interference. This article presents an innovative approach that combines enhanced Wasserstein generative adversarial network with gradient penalty (WGAN-GP) with frequency-domain constraints and cross-modal feature fusion to address these challenges. The WGAN-GP model, enhanced with spectrum-aware loss, generates high-fidelity reverberation signals. This method, coupled with a bright spot model for target echo simulation, creates a robust dataset that alleviates the issue of small sample sizes, which is critical in sonar systems. Furthermore, a multimodal convolutional network leveraging Gramian angular field (GAF) and Markov transition field (MTF) representations is introduced for effective feature extraction and classification. The proposed system achieves superior classification accuracy, averaging 94.8% at 10 dB signal-to-reverberation ratio (SRR), outperforming traditional models. Experimental validation in diverse marine environments, including the South China Sea, demonstrates the framework’s high generalization ability and strong performance under varying SRR conditions. These results establish the proposed method as a powerful tool for robust, real-time sonar target classification in challenging environments.