Abstract: Automatic Modulation Recognition (AMR) for Underwater Acoustic Communication (UWAC) signals faces severe challenges due to the strong multipath effect, Doppler spread, and ambient noise interference inherent in underwater acoustic channels, as well as the limitations of single-modal feature representations. Traditional UWAC modulation recognition methods—typically based on likelihood ratio tests or statistical feature analysis—often suffer from suboptimal performance in complex underwater environments. To address these issues, this thesis proposes a Dual-Stream Cross-Attention Network built upon dual Residual Neural Networks (ResNets) and a cross-attention mechanism. The model transforms received signals into Gramian Angular Field (GAF) images and spectrograms as inputs to the two ResNet streams, enabling cross-modal, multi-scale feature extraction. It employs cross-attention for adaptive feature fusion and introduces a feature decoupling loss function to enhance discriminative capability, thereby establishing a new paradigm for efficient UWAC signal reception. Experimental results, evaluated on real-world marine observation data, demonstrate that the proposed model achieves 90% recognition accuracy at a Signal-to-Noise Ratio (SNR) of −3 dB, with an average test-set accuracy of 91.53%.