VB-UREF: uncertainty-aware recursive evidence fusion for underwater acoustic target recognition

Robust underwater acoustic target recognition remains challenging because multi-feature evidence can be affected by low signal-to-noise ratios, non-stationary marine noise, inter-branch conflict, and locally unreliable observations. This paper proposes a variational Bayesian uncertainty-aware recursive evidence fusion framework, termed VB-UREF. The method represents the class-probability outputs of MFCC, GFCC, and time–frequency-spectrogram recognition branches in centered log-ratio coordinates and models them as noisy observations of a latent discriminative state. A Student-t Gaussian scale-mixture likelihood and variational Bayesian covariance estimation provide residual-adaptive observation precision, while Jensen-Shannon divergence and predictive entropy characterize cross-branch conflict and self-uncertainty. The resulting adaptive evidence is combined with the previous within-recording posterior through a temporally regularized logarithmic opinion pool. In the clean comparative setting, VB-UREF achieved the highest observed accuracy of 94.31% and a Macro-F1 of 94.16%. WA-DS obtained 93.48% accuracy and 93.25% Macro-F1; the 0.83-percentage-point accuracy difference was significant in a paired contiguous-block bootstrap test (p = 0.0076). Under the defined DeepShip-derived waveform-noise protocol, VB-UREF achieved 64.34%, 82.88%, 92.00%, 93.89%, 94.04%, and 94.06% accuracy from −30 to −5 dB, respectively; the strongest baseline results were 46.88% at −30 dB and 92.90% at −5 dB. Abnormal-evidence diagnostics further show that suppression arises from the complementary action of conflict, entropy, and residual-covariance terms. These results support VB-UREF as an interpretable decision-fusion method for uncertain underwater acoustic observations.