Optimal feature selection and model explanation for reef fish sound classification.

Area of Science:

• Marine Biology
• Bioacoustics
• Artificial Intelligence

Background:

• Fish sounds are crucial acoustic cues in reef ecosystems, influencing ecological processes.
• Artificial intelligence (AI) is increasingly used for detecting, classifying, and identifying fish vocalizations.
• Understanding fish sounds aids in comprehending diel behaviors and ecological roles in reefs.

Purpose of the Study:

• To classify unknown fish sounds from a subtropical rocky reef using AI.
• To evaluate the effectiveness of different feature sets, data augmentation, and explainable AI tools.
• To identify key acoustic features contributing to fish sound classification.

Main Methods:

• Supervised learning algorithms (naive Bayes, random forest, decision trees, multilayer perceptron) were employed.
• Multiclass classification was performed on four distinct classes of fish pulsed sounds.
• Data augmentation and explainable AI techniques were utilized to enhance model performance and interpretability.

Main Results:

• The proposed AI models demonstrated excellent classification performance, with a multilayer perceptron achieving 98.1% accuracy.
• Data augmentation significantly improved classification accuracy.
• Explainable AI successfully identified the specific acoustic features predictive of each sound class.

Conclusions:

• AI, particularly multilayer perceptron with data augmentation, is highly effective for classifying fish sounds in reef environments.
• Explainable AI provides valuable insights into the acoustic characteristics differentiating fish sound classes.
• Accurate fish sound recognition is essential for monitoring reef ecology and conservation efforts.