SCP-TriCA fuses 2D/1D signals for 94.86% sea-trial accuracy

A team of researchers led by Ronglai Qian has released a novel framework for underwater acoustic modulation recognition, tackling the critical challenge of distribution shift—where models trained in controlled lab conditions fail in real-world environments. They present two contributions: UAMR-ShiftBench, the first benchmark that systematically evaluates in-distribution, low-SNR, unseen-environment, unseen-communication-parameter, and measured sea-trial scenarios under a single protocol, and SCP-TriCA, a hierarchical fusion architecture that integrates heterogeneous signal representations.

SCP-TriCA fuses three modalities: two 2D representations (STFT spectrograms and cyclostationary maps) are first aligned via bidirectional cross-attention, then a 1D statistical descriptor (second- and fourth-order power spectra) is incorporated through a sample-adaptive gating mechanism. On UAMR-ShiftBench, SCP-TriCA achieves 95.33% in-distribution accuracy and 74.59% simulated OOD average, beating the strongest baseline by 5.12 percentage points. More strikingly, on two independent sea-trial subsets collected in March and November in the South China Sea, SCP-TriCA reaches 91.14% and 94.86% accuracy, exceeding baselines by 15.71 and 23.00 percentage points, respectively. This demonstrates that robust fusion of complementary modalities, especially under real-world noise and parameter shifts, can dramatically improve deployment readiness for underwater acoustic AI systems.