Underwater surveillance is vital for naval defense, marine research, and environmental monitoring. One key method relies on identifying ship radiated noise (SRN)—acoustic signals produced by engines, propellers, and onboard machinery. But recognizing these sounds in the chaotic underwater environment is challenging. Noise from marine life, shifting ocean currents, and overlapping ship traffic often drowns out critical details.
A new study by Di Zeng and colleagues (2025) introduces FCResNet5, a specialized deep learning model that dramatically improves ship noise classification by focusing on essential frequencies and optimizing data channels. This approach could transform real-time marine detection systems while slashing computational costs.
What Makes FCResNet5 Different?
Most deep learning models adapted for underwater acoustics borrow from image processing techniques, using RGB channels and broad frequency bands. This often leads to wasted computation and lower accuracy. FCResNet5 changes the game by introducing three major innovations:
- Frequency Focus: It targets frequencies below 2 kHz, where ship noise signatures are most distinct, filtering out irrelevant high-frequency clutter.
- Frequency Channelization: Instead of RGB channels, the model treats each frequency band as a channel, aligning network structure with acoustic physics rather than visual imagery.
- Model Compression: By using fewer layers and a descending channel structure, FCResNet5 delivers high accuracy with over 90% fewer FLOPs (floating-point operations) than traditional ResNet models.
Key Findings
- Improved Accuracy: FCResNet5 outperformed other advanced models on CQT and Gamma-tone spectrograms, achieving superior classification of oil tankers, cargo ships, passenger vessels, and tugboats.
- Efficiency Gains: It achieved this with dramatically reduced computational requirements—ideal for low-power underwater devices.
- Noise Robustness: The model maintained strong performance in moderate noise environments (0–10 dB SNR), making it suitable for real-world marine conditions.
- Scalability: Its lightweight design allows deployment on compact hardware like autonomous underwater vehicles or fixed hydrophone arrays.
Why It Matters
Efficient SRN recognition has huge implications:
- Naval Defense: Enables real-time tracking of potentially hostile vessels.
- Maritime Traffic Monitoring: Improves safety by identifying ship types in congested waterways.
- Environmental Research: Helps study noise pollution’s impact on marine life.
- Resource Exploration: Assists in locating and identifying support vessels in offshore drilling or wind farm operations.
What’s Next?
The authors plan to refine FCResNet5 for extreme noise conditions (below −5 dB) and explore automated compression techniques for even greater efficiency. This could pave the way for next-generation, real-time underwater detection networks.
Check out the cool NewsWade YouTube video about this article!
Article derived from: Zeng, D., Yan, S., Yang, J. et al. An efficient deep learning approach with frequency and channel optimization for underwater acoustic target recognition. Sci Rep 15, 27369 (2025). https://doi.org/10.1038/s41598-025-12452-2
