Zero-shot classification of small target on sea bottom using model-agnostic meta-learning

Area of Science:

• Marine Geophysics
• Machine Learning
• Signal Processing

Background:

• Accurate identification of small targets on the seafloor is crucial for marine operations.
• Existing classifiers struggle with out-of-distribution data, limiting their effectiveness in varied ocean environments.
• Model-agnostic meta-learning (MAML) offers a potential solution for improved generalization.

Purpose of the Study:

• To develop and evaluate a MAML-based active target classifier for identifying small seafloor targets.
• To enhance the classifier's ability to generalize to out-of-distribution samples and diverse ocean conditions.
• To compare the performance of the MAML-based classifier against conventional neural networks.

Main Methods:

• Utilized frequency-domain target and clutter scattering signals.
• Trained the classifier on data from varied bottom types (silt/clay) and incident angles (low/moderate/high).
• Applied MAML to out-of-distribution samples for improved classification of deviating targets.

Main Results:

• The MAML-based classifier demonstrated significantly superior performance compared to conventional neural networks during testing.
• The enhanced generalization capabilities of MAML were evident in its ability to classify targets in unseen ocean environments.
• Analysis of the loss landscape revealed a smooth, convex curve, explaining MAML's improved generalization.

Conclusions:

• MAML provides a robust framework for developing active target classifiers with superior generalization capabilities for seafloor object detection.
• The proposed MAML-based approach is effective in handling out-of-distribution samples, crucial for real-world oceanographic applications.
• This method offers a promising advancement for autonomous underwater vehicle (AUV) mine detection systems.