Propagation-physics-guided channel augmentation and temporal–spectral fusion for domain-robust underwater acoustic target recognition

Underwater acoustic target recognition (UATR) is essential for ocean-intelligence sensing, but deep learning-based models often suffer from performance degradation under ocean acoustic domain shift. This problem arises because ship-radiated noise recorded under different propagation channels, vessel operating states, background noise levels, and recording configurations may exhibit substantial distribution discrepancies. To address this problem, we propose a Propagation-Physics-Guided Channel Augmentation and Temporal–Spectral Fusion framework (PropCA-TS) for domain-robust UATR. Specifically, coherent multipath channel impulse responses are synthesized using an acoustic propagation simulator and convolved with received ship-radiated recordings to introduce physically interpretable channel perturbations. To complement the propagation-guided augmentation, a temporal-spectral representation is coupled with a lightweight multi-scale recognition backbone to capture complementary acoustic information while maintaining computational efficiency. Experiments on the DeepShip dataset are conducted under the strict recording-ID-based split. Under five-fold recording-level validation, PropCA-TS with ONC-consistent channel augmentation achieves 80.60% accuracy and 80.37% F1-score with only 2.31M parameters and 2.328G FLOPs. Moreover, it obtains 79.23% recording-macro accuracy and reduces the domain generalization gap and relative performance drop to 18.86 percentage points and 19.23%, respectively, demonstrating stronger robustness to unseen recording-level domain shift.