Passive Acoustic Classification of Marine Vessels under the Influence of Range, Sea States and Seasonal Changes in Sound Speed Profiles

Automatic classification of marine vessels using radiated acoustic noise is a key requirement for naval operations andunderwater surveillance. However, classifier performance degrades under varying environmental and operationalconditions such as sea state, sound speed profile (SSP), and target range. This study examines the impact of sea state variability, SSP conditions, and target range on the performance andgeneralization capability of convolutional neural networks (CNN) based marine vessel classifiers. Comprehensive mathematical models were developed to simulate marine vessel noise, sea state–induced ambientnoise, and underwater acoustic propagation. Vessel noise was generated for multiple operational speeds, whilepropagation effects were modeled using the Bellhop ray tracing model with Gaussian beam tracing, incorporating SSP(Sound Speed Profile) variability and range-dependent propagation loss. Three CNN models were trained using datasetscorresponding to sea states one, four, and seven, and evaluated across all environmental and operational scenarios. The classification performance varied significantly with environmental and operational conditions. The CNN trained onsea state seven exhibited superior generalization across other sea states. A performance drop of approximately 30% wasobserved when the target range reached 25 km. SSP dependent analysis revealed strong sensitivity in convergence andshadow zones, where correct classification depended critically on precise range information. ROC analysis showed thatmodels trained under rough sea conditions achieved about 20% better generalization compared to those trained on lowersea states. The results demonstrate that environmental and operational variability strongly influences acoustic classificationperformance. Training with data representing rough sea conditions improves model robustness and generalization. Thestudy highlights the necessity of accurately modeled synthetic datasets as a practical alternative to limited and costly real-world acoustic recordings with incomplete environmental characterization.