Insights and approaches using deep learning to classify wildlife.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Wildlife Conservation

Background:

• Automated identification of wildlife using intelligent software is crucial for conservation and management.
• Deep learning methods, particularly convolutional neural networks (CNNs), are state-of-the-art for image classification but can be opaque to non-experts.

Purpose of the Study:

• To elucidate the methods and extracted features used by CNNs for classifying wildlife species from camera-trap data.
• To improve the interpretability of CNN models in the context of wildlife identification.

Main Methods:

• Training a CNN on a dataset of 111,467 camera-trap images to classify 20 African wildlife species.
• Applying gradient-weighted class-activation-mapping (Grad-CAM) to visualize salient image features used by the CNN.
• Utilizing mutual information to identify key neurons and hierarchical clustering for feature vector analysis.

Main Results:

• Achieved an overall classification accuracy of 87.5% for 20 African wildlife species.
• Grad-CAM highlighted features similar to those humans use for species identification.
• Identified dataset biases and evaluated model performance on known versus unknown species.

Conclusions:

• CNNs offer a powerful tool for automated wildlife species classification from camera-trap data.
• Interpretable AI methods like Grad-CAM enhance understanding of model decision-making.
• The approach aids in wildlife management and reveals insights into model behavior and data characteristics.