A multimodal framework for long-range vessel classification using optical images and radar full-pulse sequences

In contemporary maritime situational-awareness systems, conventional electro-optical sensors are acutely vulnerable to atmospheric scattering and sea-clutter perturbations, precipitating a pronounced degradation in long-range vessel recognition fidelity. Consequently, monomodal imaging-centric paradigms fail to satisfy operational exigencies under complex sea-state conditions, thereby constraining the efficacy of maritime security early-warning frameworks. To redress this deficiency, the present study curates a multimodal corpus comprising thirty-six distinct vessel taxa and introduces a novel visual-signal dual-Modality recognition model that synergistically couples high-resolution visible-light imagery with radar full-pulse signatures. Specifically, a Vision Transformer-B/16 backbone distils hierarchical visual descriptors, while a Bidirectional Long Short-Term Memory network encodes the temporal dynamics inherent in radar pulse streams; these representations are integrated via a dynamic feature-fusion mechanism and subsequently reconciled through an adaptive decision-level ensemble for robust classification. Empirical evaluation evidences that the proposed methodology attains a mean recognition accuracy of 86.3 % on the held-out test partition—an absolute gain of 15.5 % relative to monomodal baselines. Both analytical derivations and experimental corroborations affirm that cross-modal complementarity materially enhances environmental robustness, thereby furnishing pivotal technical underpinnings for next-generation intelligent maritime surveillance architectures. These contributions confer substantive theoretical novelty and demonstrable engineering utility across applications spanning maritime situational awareness, maritime search-and-rescue, and waterway monitoring domains.