Automatic Classification of Cichlid Behaviors Using 3D Convolutional Residual Networks.

Area of Science:

• Ethology
• Machine Learning
• Animal Behavior

Background:

• Understanding complex animal behaviors is crucial for survival and reproduction.
• Large video datasets offer opportunities but require automated quantification tools.
• Behavioral analysis in fish, particularly cichlids, can provide insights into social and survival strategies.

Purpose of the Study:

• To develop and validate a machine learning model for automated classification of complex fish behaviors.
• To quantify specific behaviors such as spitting, scooping, fin swipes, and spawning in Lake Malawi cichlid fishes.
• To determine if animal intent (e.g., feeding vs. bower construction) can be inferred from classified behaviors.

Main Methods:

• Utilized pixel-based hidden Markov modeling and density-based spatiotemporal clustering to identify sand disturbance events.
• Applied a 3D Residual Network (ResNet) trained on over 11,000 manually annotated video clips.
• Classified behaviors into 10 distinct categories, including spitting, scooping, fin swipes, and spawning.

Main Results:

• The 3D ResNet achieved over 76% accuracy in classifying fish behaviors.
• Successfully distinguished between different types of sand disturbance events.
• Differentiated between spits and scoops related to bower construction versus feeding, indicating intent.

Conclusions:

• 3D Residual Networks are effective tools for automatically quantifying complex behaviors in fish from large video datasets.
• Automated behavioral classification can provide nuanced insights into animal intent and context-specific actions.
• This approach facilitates large-scale ethological research on survival and reproductive behaviors.