Underwater source localization via active deep complex residual network in a shallow-water waveguide.

Area of Science:

• Acoustics
• Signal Processing
• Machine Learning

Background:

• Current deep learning for underwater source localization often uses real-valued data, potentially underutilizing information.
• High labeled data requirements and poor adaptability limit deep learning applications in diverse underwater environments.

Purpose of the Study:

• To develop an efficient and adaptable deep learning framework for underwater source localization.
• To address limitations of real-valued representations and extensive labeling in current methods.

Main Methods:

• Proposed an end-to-end active learning framework (ADCRN) with a deep complex residual network (DCRN-CCL) backbone.
• Integrated an adaptively weighted uncertainty-diversity query strategy and model-based transfer learning.
• Employed complex-domain circle loss and mean-based probabilistic fusion for enhanced feature discrimination.

Main Results:

• Achieved approximately 12% higher accuracy and reduced mean absolute deviation by 1.2 compared to real-valued deep learning methods.
• Maintained performance comparable to DCRN-CCL on the SWellEx-96 dataset.
• Required only 37.89% of total samples for labeling and 69.87% of training time.

Conclusions:

• The proposed ADCRN framework effectively enhances underwater source localization accuracy and efficiency.
• Complex-valued feature representation and active learning significantly reduce data dependency and improve generalization.