Automatic Identification of Pangolin Behavior Using Deep Learning Based on Temporal Relative Attention Mechanism

Area of Science:

• Zoology and Animal Behavior
• Computer Science and Artificial Intelligence
• Conservation Biology

Background:

• Pangolin populations are declining, making captive breeding crucial for species survival.
• Current artificial breeding success is limited by a poor understanding of pangolin reproductive behaviors.
• Machine vision offers non-invasive, continuous monitoring to reduce animal stress and gather behavioral data.

Purpose of the Study:

• To develop and evaluate a deep learning model for monitoring and recognizing pangolin behaviors, including breeding and daily activities.
• To establish a temporal relation and attention mechanism network (PBATn) for enhanced behavioral analysis.
• To provide a tool that supports conservation efforts through improved understanding of pangolin reproduction.

Main Methods:

• A dataset of 11,476 videos featuring pangolin breeding and daily behaviors was utilized.
• The proposed Pangolin breeding attention and transfer network (PBATn) was trained and validated.
• The PBATn model was tested against established methods like SlowFast, X3D, TANet, and TSN using metrics such as mAP, accuracy, recall, specificity, and F1 score.

Main Results:

• The PBATn model achieved high accuracy (98.95% training, 96.11% validation) and low loss values.
• PBATn demonstrated superior performance over baseline models on the test set, with mAP of 97.50% and average accuracy of 99.17%.
• Specific breeding behaviors like chasing and mounting were recognized with high accuracy (94.00% and 98.50%, respectively).

Conclusions:

• The PBATn deep learning system accurately monitors pangolin breeding behaviors, outperforming existing methods.
• This technology can significantly aid in analyzing pangolin behavior, contributing to conservation strategies.
• The system's effectiveness supports the potential of AI in wildlife conservation and captive breeding programs.