Interpretable deep dictionary learning for sound speed profiles with uncertainties.

Area of Science:

• Oceanography
• Signal Processing
• Machine Learning

Background:

• Uncertainties in ocean sound speed profiles (SSPs) hinder underwater applications and knowledge acquisition.
• Existing ocean observing systems face challenges in accurately measuring and estimating SSPs.
• Measurement errors (Gaussian noise) and dynamic disturbances affect SSP accuracy.

Purpose of the Study:

• To develop an interpretable deep dictionary learning model for processing uncertain SSPs.
• To reduce SSP uncertainties and gain insights into ocean processes.
• To effectively mitigate noise and detect/localize SSP disturbances.

Main Methods:

• Proposed an interpretable deep dictionary learning model.
• Unrolled the K-singular value decomposition algorithm into a neural network.
• Trained the neural network using supervised learning with carefully designed data and initializations.

Main Results:

• The proposed method demonstrated superior performance compared to classical baselines.
• Effectively mitigated noise corruptions in SSP data.
• Successfully detected and localized SSP disturbances and anomalies.

Conclusions:

• The interpretable deep dictionary learning model offers a robust solution for uncertain SSP processing.
• The method enhances the reliability of oceanographic data for downstream applications.
• This approach provides a pathway for better understanding ocean dynamics through improved SSP analysis.