Hybrid Swin vision transformer with IMFCC for enhanced underwater acoustic noise reduction

Abstract

Underwater wireless communication is essential for deep-sea exploration, environmental monitoring, and Autonomous Underwater Vehicle (AUV) coordination. Its performance is severely affected by ambient noise, multipath propagation, reverberation, and non-stationary interference, which degrade acoustic signal quality. Existing noise reduction techniques, such as Wiener filtering, spectral subtraction, and wavelet denoising, struggle to adapt to rapidly changing underwater conditions. Centralized Deep Learning (DL) based enhancement models are also unsuitable for distributed Underwater Wireless Sensor Networks (UWSNs) due to privacy, bandwidth, and real-time constraints. To overcome these challenges, this work proposes a Federated Learning–Driven Hybrid Swin Vision Transformer with Improved MFCCs (Fed-SVTIMFCC) for intelligent underwater acoustic noise reduction. Federated learning enables multiple underwater nodes to collaboratively train a denoising model without sharing raw acoustic data, ensuring data privacy and adaptability. The Swin Vision Transformer (SVT) captures local spectral cues and long-range dependencies to enhance reconstruction, while IMFCCs provide robust features for distinguishing noise from useful signals. Experimental results on real underwater datasets demonstrate significant improvements in SNR (26.94 dB), PSNR (33.28 dB), and SSIM (0.947), establishing Fed-SVT-IMFCC as a powerful solution for next-generation underwater communication systems.