Multifrequency species classification of acoustic-trawl survey data using semi-supervised learning with class...

Area of Science:

• Marine ecology
• Acoustic sensing technology
• Statistical modeling

Background:

• Acoustic surveys commonly use multifrequency backscatter to assess fish and plankton populations.
• Direct sampling is crucial for validating acoustic data but can be sparse or unavailable.
• Accurate species classification is essential for reliable abundance estimates.

Purpose of the Study:

• To develop a generalized Gaussian mixture model for classifying multifrequency acoustic backscatter.
• To enable species classification even when not all species are known beforehand (semi-supervised learning).
• To apply this novel classification method to ecological data from the eastern Bering Sea.

Main Methods:

• Development of a generalized Gaussian mixture model incorporating semi-supervised learning and class discovery.
• Application of the model to acoustic backscatter data collected in the eastern Bering Sea.
• Analysis of data from summer surveys conducted in 2004, 2007, and 2008.

Main Results:

• The generalized Gaussian mixture model successfully classified multifrequency acoustic backscatter.
• The model identified key biological classes, including walleye pollock and euphausiids.
• Two additional major species classes in the upper water column were also identified.

Conclusions:

• The developed semi-supervised model effectively classifies acoustic backscatter data without requiring complete prior species knowledge.
• This approach enhances the utility of acoustic surveys for estimating fish and plankton abundance, particularly in data-limited situations.
• The study successfully identified ecologically important species in the eastern Bering Sea, contributing to marine ecosystem understanding.