Abstract: Modulation recognition of underwater acoustic communication signals is a key task in signal interception processing. In recent years, recognition methods combining time-frequency features with neural networks have attracted considerable attention. To address the insufficient generalization ability of conventional modulation recognition methods under cross-noise-distribution conditions and the confusion between similar modulation classes caused by single-feature representation, an underwater acoustic communication modulation recognition method based on a dual-branch neural network is proposed in this paper. Taking the raw signal waveform as input, the proposed method constructs a pre-trained audio encoder branch and a spectral feature branch, and further designs an adaptive decision fusion strategy guided by static fusion to jointly optimize the decisions of the two branches. Experimental results show that, under test conditions involving unseen noise types, the proposed method improves the average recognition accuracy by 13% compared with the best-performing traditional spectrogram-based method. Meanwhile, the discrimination ability for similar modulation classes is enhanced, with the proportion of QPSK samples misclassified as BPSK reduced from 18.4% to 5.0%, effectively alleviating confusion among similar categories.