Adaptive Spectral Line Enhancement Using Deep Feature Extraction Guided FISTA-Net for Noise-Resilient Underwater Passive Sonar Detection and Classification

Abstract

The low frequency analysis and recording (LOFAR) spectrum plays a vital role in underwater passive sonar target detection. High noise levels in the marine environment degrade the quality of spectral lines in LOFAR spectrograms and hinder target detection. This article proposes a novel hybrid approach for LOFAR spectrum enhancement that integrates a sparse optimization algorithm and a deep learning approach for adaptive, noise resilient spectral line preservation. The proposed architecture employs U-shaped network (U-Net) as a spectral line selector and a cascaded fast iterative shrinkage-thresholding algorithm (FISTA) network for adaptive noise suppression. The U-Net selectively emphasizes salient frequency components, guiding the subsequent FISTA blocks to perform enhancement while preserving critical spectral lines. A signal-to-noise ratio (SNR) estimation module dynamically evaluates the noise level and adaptively determines the appropriate denoising level, enabling precise noise suppression. Experimental results using ShipsEar data sets demonstrate that the proposed method, namely, deep feature extraction guided FISTA-Net achieves an SNR gain of 3.37 dB over the state-of-the-art method, namely, fast block processed sparsity adaptive line enhancer (ALE) and 7.12 dB over sparsity induced ALE. In addition to the SNR improvement, the proposed method achieves an average signal-to-distortion ratio of 21.64 dB, which is 2.67 dB higher than the best performing existing method, further confirming its effectiveness in preserving signal quality while suppressing noise. For the evaluation of classification performance, a common residual neural network classifier was employed across all methods. The proposed approach achieves a classification accuracy of 96.95%, outperforming existing state-of-the-art techniques, confirming the proposed method's effectiveness in enhancing the detection and recognition of underwater acoustic targets.