PBGMR: A Physics-Informed Bayesian Graph Neural Network for Robust Multimodal Underwater Target Recognition

Existing approaches for underwater autonomous target recognition (UATR) often suffer from limited accuracy and robustness due to complex underwater environments, noise interference during measurements, and inadequate utilization of multimodal sonar data. To overcome these limitations, we propose a novel physics-constrained graph neural network for robust multimodal UATR (PBGMR). The proposed method uniquely integrates physics-informed feature extraction, a physically consistent graph construction mechanism, and multimodal Bayesian fusion within a unified framework. Specifically, we extract robust time–frequency and spatial features from mid-frequency active sonar echoes using Fourier-based synchrosqueezing transform (FSST) and beam–time–range (BTR) mapping, while high-level spatial features are derived from forward-looking sonar images via a ConvNeXt-V2 extractor. These features are embedded as nodes in a graph whose edges reflect physically plausible associations. A residual multi-head graph convolutional network is then proposed for intra- and intermodal feature propagation and fusion. Bayesian regularization employing variational inference and an adaptive multi-objective loss function further enhances measurement robustness. Extensive experiments on a comprehensive sensor dataset collected from lake trials demonstrate that the proposed PBGMR achieves superior recognition performance, significantly outperforming single-modality baselines by up to 14% and existing state-of-the-art methods, with an accuracy exceeding 93%.