Redefining Channel Estimation in Underwater Acoustic OFDM Systems with Deep Neural Network

This paper introduces a novel method for channel estimation in underwater acoustic communication in an autonomous underwater vehicular network. The proposed method employs a denoising technique to refine least squares (LS) channel estimates using deep image prior (DIP). By establishing an equivalence between underwater acoustic (UWA) channel estimation and image denoising, we leverage DIP to enhance estimation accuracy. The proposed approach is validated on the Norway continental shelf (NCS1) watermark dataset, demonstrating superior performance with average mean square error reductions of 96.64% and 96.09% compared to LS and the deep denoising convolutional neural network (DnCNN), respectively. Furthermore, the proposed analysis of pilot symbol utilization in the DIP-based estimator shows a 46.47 % error reduction, even when using only 25 % of the pilot symbols. By efficiently utilizing available resources, the proposed method enhances spectral efficiency and enables accurate estimation, even with limited pilot signals.