A Study on the Detection of Cattle in UAV Images Using Deep Learning.

Area of Science:

• Agricultural Technology
• Computer Vision
• Animal Science

Background:

• Extensive cattle farming presents challenges for monitoring due to large areas and dispersed animals.
• Advancements in Unmanned Aerial Vehicles (UAVs), imaging, and deep learning, specifically Convolutional Neural Networks (CNNs), offer potential solutions for improved cattle management.
• Current research on UAVs for cattle monitoring has significant gaps, necessitating further investigation.

Purpose of the Study:

• To achieve high accuracy in detecting Canchim breed cattle (similar to Nelore) using aerial imagery.
• To identify the optimal Ground Sample Distance (GSD) for effective cattle detection.
• To determine the most suitable CNN architecture for cattle detection from UAV images.

Main Methods:

• Trained 900 models using 15 different CNN architectures across 3 spatial resolutions and 2 datasets.
• Utilized 1853 aerial images containing 8629 cattle samples for training and validation.
• Employed 10-fold cross-validation for robust model evaluation.

Main Results:

• Several CNN architectures demonstrated robustness in detecting cattle, even in suboptimal imaging conditions.
• The study analyzed various factors influencing cattle detection accuracy from UAV-captured aerial images.
• Viability of UAVs for effective cattle monitoring was indicated by the reliable detection capabilities of tested models.

Conclusions:

• UAVs equipped with deep learning offer a viable solution for monitoring cattle in extensive farming operations.
• The developed methods show potential for improving herd management and animal welfare through automated detection.
• Further research can build upon these findings to refine UAV-based cattle monitoring systems.