UAV-Acquired Dataset for Farm Intrusion Detection

Description

This dataset was developed at Soroti University (Uganda) to support machine learning research in UAV-based farm security. It contains 2,067 RGB images organized into three categories:

Animals (907 images) – livestock such as goats and cows, captured under varying poses and lighting conditions.

People (588 images) – potential human intruders recorded primarily from UAV aerial perspectives.

Empty Spaces (572 images) – unoccupied farmland areas (grass, soil, fences) included as a negative class to reduce false positives.

Key features:

Acquired using a DJI Mavic 3 Cine UAV and supplemented with smartphone ground captures.

Images resized to 255×255 pixels for computational efficiency while retaining distinguishing features.

Collected across diverse altitudes, times of day, and environmental conditions.

Accompanied by baseline benchmarking results using CNN and MobileNetV3 Small classifiers (96.8% and 98.4% accuracy, respectively).

Ethically curated: personally identifiable features (faces, license plates) were blurred or removed; data collected with landowner consent.

This dataset represents the first publicly available UAV-acquired intrusion detection dataset tailored to agricultural security. It enables benchmarking, reproducibility, and development of AI-driven solutions for livestock farm protection.