Wavelet-Enhanced DEMON Processing and Deep Learning for Passive SONAR Target Recognition

Abstract

This paper presents an integrated underwater Automatic Target Recognition (ATR) framework that synergizes advanced signal processing with deep learning. To eliminate non-stationary ambient ocean noise and isolate low-frequency modulations, the raw acoustic pressure signals undergo pre-denoising via Discrete Wavelet Transform (DWT with db4 mother wavelet and soft-thresholding), followed by Detection of Envelope Modulation on Noise (DEMON) analysis to yield 128-bin spectral feature vectors. Classification is performed using a dedicated 7-layer 1D Convolutional Neural Network (1D-CNN) architecture terminating in a Softmax output layer. Acoustic simulations in MATLAB R2024b generated a balanced dataset of 1,200 audio samples (200 samples per class across six distinct targets: fast boat, rubber zodiac, human diver, giant turtle, dolphin, and mini submarine) propagated through a 20–40 path shallow-water multipath channel ( ). Experimental evaluations demonstrate an overall classification accuracy of 94.93% and a macro-averaged F1-score of 0.9505. Coupled with an average inference latency of 11.40 ms per target and Euclidean distance-based thresholding for out-of-distribution anomaly detection, the proposed system proves computationally efficient and well-suited for real-time passive underwater surveillance.