A Passive Detection Method for Underwater Weak Targets by Integrating Similarity Networks with Graph Attention Neural Networks

To address the challenges of underwater weak targets' radiation signals being easily overwhelmed by environmental noise and the significant degradation of traditional detection methods under low signal-to-noise ratios and cross-scenario conditions, this paper proposes CORR-GAT, a graph-attention-based passive detection method for weak underwater targets built on a phase-space graph representation. For comparative purposes, three reference baselines are employed: narrowband energy detection as the traditional detection baseline, CORR-NET as the traditional similarity network detection baseline, and CORR-GCN as the graph convolutional end-to-end learning baseline. Simulation and sea-trial results demonstrate that at a false alarm probability of 1%, CORR-GAT achieves an 80% detection probability with an SNR approximately 3.5 dB lower than narrowband energy detection, and reduces the minimum detectable SNR by approximately 3.76 dB. Overall detection performance surpasses both CORR-GCN and CORR-NET, demonstrating superior low-SNR detection capability and cross-scenario generalization. This work contributes a unified detection framework integrating 'similarity-network modeling with graph attention learning,' reducing reliance on prior noise models and manual statistics through learnable attention aggregation.