Using tri-axial accelerometer loggers to identify spawning behaviours of large pelagic fish

Area of Science:

• Animal behavior analysis
• Machine learning applications
• Bio-logging technology

Background:

• Tri-axial accelerometers enable remote monitoring of animal behavior, but large datasets require semi-automated analysis.
• Marine fish exhibit complex burst behaviors that are challenging to differentiate using traditional methods.
• Accurate automated techniques for identifying natural fish behaviors are limited, especially where direct observation is impossible.

Purpose of the Study:

• To develop and validate a machine learning model for classifying marine fish behaviors using accelerometer data.
• To differentiate between various burst behaviors in fish, including feeding, courtship, and escape responses.
• To apply the developed model to free-ranging fish to identify behaviors in their natural environment.

Main Methods:

• A random forest machine learning algorithm was trained on 624 hours of accelerometer data from captive yellowtail kingfish.
• Five distinct behaviors (swim, feed, chafe, escape, courtship) were identified and used to train the model with 58 predictive variables.
• The model's performance was evaluated using accuracy and F1 scores for each behavioral class.

Main Results:

• The machine learning model achieved an overall accuracy of 94% in classifying fish behaviors.
• Classification accuracy varied by behavior, with F1 scores ranging from 0.48 for 'chafe' to 0.99 for 'swim'.
• The model successfully predicted all five defined behaviors in free-ranging kingfish, identifying 19 courtship events.

Conclusions:

• Supervised machine learning offers a novel approach for analyzing behavior in free-ranging animals from accelerometer data.
• This method overcomes limitations of direct observation, enabling prediction from unseen datasets.
• Ambiguous spawning and courtship behaviors of large pelagic fish were identified in their natural context.