Automatically identifying, counting, and describing wild animals in camera-trap images with deep learning

Area of Science:

• Ecology
• Wildlife Biology
• Artificial Intelligence

Background:

• Accurate wildlife data is crucial for ecosystem study and conservation.
• Motion-sensor camera traps offer inexpensive, unobtrusive wildlife data collection.
• Manual data extraction from camera trap images is time-consuming and costly.

Purpose of the Study:

• To investigate the automatic, accurate, and inexpensive collection of wildlife data using AI.
• To develop deep learning models for identifying, counting, and describing animal behaviors from camera trap images.

Main Methods:

• Training deep convolutional neural networks on the 3.2 million-image Snapshot Serengeti dataset.
• Utilizing deep learning to identify, count, and describe behaviors of 48 species.
• Implementing a confidence threshold for automated classification.

Main Results:

• Deep neural networks achieved >93.8% accuracy in automatic animal identification.
• Automated identification covered 99.3% of the dataset with 96.6% accuracy, matching human volunteers.
• The system saved an estimated 8.4 years of human labeling effort.

Conclusions:

• Deep learning significantly enhances the efficiency of camera trap data analysis.
• AI-driven automation can overcome data extraction roadblocks for wildlife research.
• This technology facilitates high-volume, real-time wildlife data collection for ecological studies.