Joint deconvolution and classification with applications to passive acoustic underwater multipath.

Area of Science:

• Signal Processing
• Machine Learning
• Bioacoustics

Background:

• Classifying signals corrupted by noise and unknown linear time-invariant (LTI) filtering, such as multipath, is challenging.
• Traditional methods often struggle with performance degradation under such conditions.

Purpose of the Study:

• To develop a robust signal classification method that accounts for unknown LTI filtering and noise.
• To improve classification accuracy by avoiding explicit signal deconvolution.

Main Methods:

• A maximum a posteriori (MAP) approach for joint deconvolution and classification was considered.
• A novel variant of Quadratic Discriminant Analysis (QDA) was proposed, probabilistically accounting for unknown LTI filtering.
• The QDA classifier was tested directly on signals and on subband-power features.

Main Results:

• The proposed QDA classifier significantly improved classification performance compared to traditional methods.
• Performance gains were observed across a range of signal-to-noise ratios.
• The method demonstrated effectiveness on both simulated data and real Bowhead whale vocalizations.

Conclusions:

• Jointly considering deconvolution with classification, or using a classifier that probabilistically accounts for filtering, dramatically improves performance.
• The proposed QDA classifier offers a more accurate and robust solution for classifying corrupted signals.
• This approach has potential applications in areas like bioacoustic analysis and communication systems.