A Depth Classification Method for Deep-Sea Transient Sources using YOLOv5

The depth classification for transient sources in deep-sea environments attracts much attention. Most existing depth classification methods for deep-sea underwater transient sources, including matched field processing based on Arrival time differences (MFP-ATDs), require relatively high signal-to-noise ratio (SNR). However, due to the relatively small source level and the large transmission loss (i.e., caused by the spreading loss and the bottom bouncing loss) of the transient signal, the input SNR at the receiving end is often at a low level. And the performance of existing depth classification methods degrades significantly. To improve the depth classification performance when the input SNR is relatively low, we propose a method for depth classification of transient sources using time-frequency features and YOLOv5. In the proposed method, we focus on the four paths of the bottom bouncing area in deep sea environment (i.e., bottom reflection, B; surface-bottom reflection, SB; bottom-surface reflection, BS; surface-bottom surface reflection, SBS). Specifically, we exploit the time-frequency feature of transient sources formed by the four-path channel, and use YOLOv5 to discriminate the difference of the time-frequency structures of the surface and submerged sources. To extract the useful feature of the transient signal from the time-frequency spectrogram, YOLOv5 mainly uses backbone and multi-scale feature fusion module (MSFFM). YOLOv5 can learn detailed features from various textures by backbone during the training process. Moreover, it can address the complex and variable background noise by MSFFM. Therefore, YOLOv5 can reduce the effect of complex backgrounds due to the low SNR. Simulation results show that the proposed method has superior classification performance compared to MF-ATDs on the relatively low SNR condition.