Complex data labeling with deep learning methods: Lessons from fisheries acoustics.

Area of Science:

• Marine acoustics
• Fisheries science
• Machine learning applications

Background:

• Acoustic backscattered signals are crucial for fish stock assessment and marine ecosystem monitoring.
• Vast amounts of acoustic data require time-consuming expert labeling, especially for complex echograms.
• Accurate labeling is critical for the reliability of fisheries and ecological analyses.

Purpose of the Study:

• To investigate the application of supervised learning, specifically convolutional neural networks (CNNs), for echogram labeling.
• To demonstrate how CNNs trained on non-stationary datasets can assist in identifying data requiring human expert correction.
• To explore the potential for standardizing the echogram labeling process in fisheries acoustics.

Main Methods:

• Quantitative and qualitative analysis of acoustic backscattered signals.
• Development and training of convolutional neural networks (CNNs) using non-stationary datasets.
• Application of trained CNNs to identify challenging sections of new echogram datasets.

Main Results:

• CNNs trained on non-stationary datasets can effectively highlight portions of new datasets that require expert human correction.
• The proposed method shows promise in reducing the burden of manual echogram labeling.
• This approach offers a potential pathway toward standardizing labeling in fisheries acoustics.

Conclusions:

• Supervised learning, particularly CNNs, can significantly benefit the echogram labeling process in fisheries acoustics.
• This method provides a valuable tool for improving the efficiency and consistency of data analysis.
• The study serves as a case study for addressing non-obvious data labeling challenges in scientific research.