Label distribution-guided transfer learning for underwater source localization.

Area of Science:

• Acoustics
• Machine Learning
• Signal Processing

Background:

• Deep neural networks (DNNs) for underwater source localization require extensive data and computational resources.
• Training DNNs is often hindered by the scarcity and high cost of acquiring experimental underwater acoustic data.

Purpose of the Study:

• To propose a novel transfer learning approach, label distribution-guided transfer learning (LD-TL), to address data limitations in underwater source localization.
• To enhance the performance of underwater source localization using deep learning with minimal experimental data.

Main Methods:

• A one-dimensional convolutional neural network (1D-CNN) was pre-trained using simulated underwater acoustic propagation data.
• The pre-trained 1D-CNN was fine-tuned using a limited amount of experimental data labeled with distribution vectors.
• Label distribution vectors were employed instead of traditional one-hot encoded vectors for fine-tuning.

Main Results:

• The proposed LD-TL method significantly improved the performance of underwater source localization.
• Effective localization was achieved even with a very limited amount of experimental data.
• The use of label distribution vectors proved beneficial for fine-tuning the deep learning model.

Conclusions:

• LD-TL offers an effective solution for data-scarce underwater source localization problems.
• The approach demonstrates the potential of transfer learning and label distribution learning in acoustic signal processing.
• This method reduces the reliance on large experimental datasets, making DNNs more practical for underwater applications.