Improving Animal Monitoring Using Small Unmanned Aircraft Systems (sUAS) and Deep Learning Networks.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Wildlife Ecology

Background:

• Small unmanned aircraft systems (sUAS) offer customizable, accessible, and minimally disruptive animal monitoring.
• Automated animal identification from sUAS imagery can address critical issues like wildlife-vehicle collisions, especially on airport grounds.

Purpose of the Study:

• To demonstrate automated identification of four animal species using deep learning models trained on sUAS-collected images.
• To evaluate the performance of Convolutional Neural Networks (CNN) and Deep Residual Networks (ResNet) for animal classification.

Main Methods:

• Captured 1288 images of cattle, horses, Canada Geese, and white-tailed deer using an sUAS with visible spectrum cameras.
• Developed a four-class classification problem using deep learning neural networks.
• Compared the performance of CNN and ResNet (specifically ResNet 18) models.

Main Results:

• The ResNet 18 model achieved 99.18% overall accuracy (OA) and a Kappa statistic of 0.98.
• CNN models achieved a maximum OA of 84.55% and a Kappa of 0.79.
• ResNet demonstrated effectiveness with a relatively small training dataset.

Conclusions:

• ResNet is a highly effective algorithm for distinguishing between the four tested animal species.
• Deep learning, particularly ResNet, shows significant promise for classifying larger, more diverse animal datasets from sUAS imagery.
• Automated animal classification using sUAS and deep learning can contribute to wildlife management and aviation safety.