A machine learning-based underwater noise classification method

Abstract

We proposed a machine learning-based underwater noise classification method that extracts five underwater noise features (the 1/3 octave noise spectrum level (NL), time–frequency spectrum, power spectral density (PSD), Mel-frequency cepstral coefficient, and Mel filter bank energy) from three domains (the frequency, time–frequency and Mel transform domains). We classified underwater noise using the support vector machine (SVM) and convolutional neural networks (CNN) methods and verified the results using the original data from five classes of typical underwater noise and noise-added data with different signal-to-noise ratios (SNRs). The results show that under different SNR conditions, the classification performance were better with the input features of NL and PSD; when the SNR was −10 dB, the corresponding classification accuracies were 98.95% and 97.65%, respectively. The CNN method outperformed the SVM method for classifying underwater noise, and when the SNR was above −20 dB, the mean classification accuracies of the SVM and CNN methods were 87.8% and 95.6%, respectively.