Prediction of the Low-Frequency Ambient Noise Level in the Northeast South China Sea Based on a Gated Recurrent Unit Network Model

The prediction of the oceanic ambient noise level is important for marine mammal protection and practical applications of sonar systems; however, the ambient noise level often changes dynamically as a result of atmospheric processes, oceanic processes, and human activities in the ocean. Accurate prediction of the ambient noise level is thus a challenging problem. In this letter, a data-driven ambient noise level prediction method based on a Gated Recurrent Unit network (GRU-NL) is proposed. A long-term (21-month) dataset of low-frequency oceanic ambient noise (50–1600 Hz) in the northeast South China Sea was collected. The GRU-NL model uses air–sea variables as inputs and outputs the predicted ambient noise level. Statistical results show that the proposed model achieves a satisfying prediction accuracy, with a root-mean-square error (RMSE) lower than 2.5 dB under conditions of 500, 1000, and 1600 Hz for a long-term dataset of roughly four months. The average prediction error is below −2dB, and the noise spectral level in the interval between 55 and 65 dB has the smallest prediction errors. This verifies that the proposed model has good potential in predicting the noise level of dynamically changing oceanic environments.