BiCAG-Net for marine mammal acoustic signal recognition in shipboard environmental perception

Objective The acoustic signals of marine mammals are easily masked by ship radiated noise. Existing multi-feature recognition methods lack sufficient cross-representation interaction and exhibit poor robustness under low signal-to-noise ratio (SNR) conditions. To address these issues, a two-branch interactive fusion recognition method based on bidirectional cross-attention gated fusion network (BiCAG-Net) is proposed in this paper.

Method Two complementary time-frequency representations, namely the log-Mel spectrogram and Constant-Q Transform (CQT), are constructed as dual inputs. A ResNet18 branch is employed to extract local spectral texture features from the log-Mel spectrogram, while a Conformer branch is used to extract harmonic temporal patterns from the CQT representation. A bidirectional cross-attention module is introduced to enable deep complementary information exchange between the two branches. A gated decision module based on fused embeddings, prediction probabilities, and entropy-based uncertainty estimates is designed to adaptively assign weights to different branches. A multi-task joint loss function is adopted to facilitate collaborative model optimization. A dataset containing 8805 audio clips from 9 marine mammal species is utilized, and four types of real-world vessel noise are superimposed to generate test samples at SNR levels of 20, 10, 0 and -10 dB. CRNN, ViT and PANNs are selected as baseline models for comparison.

Result s At an SNR of 20 dB, the proposed model achieves a macro-averaged F1-score of 0.968 8 and an accuracy of 0.979 6. It consistently achieves the highest Macro-F1 scores across all four SNR conditions and maintains a Macro-F1 of 0.873 8 even under severe noise interference at −10 dB SNR. Ablation studies demonstrate that dual-branch feature matching, bidirectional cross-attention and gated fusion are key structures responsible for performance improvement. With only 17.96×10⁶ parameters, the model requires merely 21.06 ms for single-sample inference on a CPU, meeting the real-time processing requirements of shipboard systems. T-SNE visualization and confusion matrix analysis verify that the model can effectively discriminate between acoustically similar species.

Conclusion BiCAG-Net exhibits strong robustness against vessel noise interference in marine mammal acoustic signal recognition. It provides technical support for shipboard passive acoustic monitoring (PAM), early warning of marine mammal risks, and ecological ship route optimization.