A data augmentation method for underwater acoustic target recognition based on data-driven and physical models

The difficulty in acquiring underwater acoustic target data often leads to insufficient sample sizes for training recognizers. To address this issue，a data augmentation method that integrates physical processes with data-driven approaches is proposed. Three key aspects are investigated：sound source signal generation，channel propagation simulation，and receiver environment simulation. A time-domain sound source generation model based on generative adversarial networks is established to simulate target radiated noise. The BELLHOP3D ocean chan⁃ nel simulation model is incorporated to simulate the physical propagation process of acoustic signals in the marine environment through convolution operations. A receiver-side data augmentation strategy is presented，which com⁃ bines factors such as hydrophone response characteristics and environmental noise. Classification experiments are conducted using the VGG16 model for validation. When the number of original training samples is 600，the models trained with individual augmentation strategies can improve the recognition accuracy by 1. 07% to 11%. The results demonstrate that the augmented data generated by this method improves classification accuracy under small sample conditions，verifying the effectiveness of the combined physical and data-driven augmentation approach.