Enhancing YOLO with Adversarial and Transfer Learning for UAV-Based Urban Wildlife Conservation

Abstract

Urban wildlife conservation has gained increasing importance globally, especially in rapidly urbanizing regions like Saudi Arabia, where habitat fragmentation and biodiversity loss are critical concerns. UAV (unmanned aerial vehicle) imagery has emerged as a promising solution for wildlife monitoring, providing high-resolution, real-time data collection even in difficult-to-access urban areas. However, the complexity of urban environments and interference from human activities pose significant challenges to the effective use of UAV imagery. While deep learning (DL) models offer robust solutions for wildlife monitoring in UAV imagery, their performance is often constrained by the limited availability of high-quality training datasets. To the best of our knowledge, this study presents the first attempt to address these challenges by harnessing the combined power of adversarial and transfer learning. The research enhances the state-of-the-art YOLOv8 (You Only Look Once, version 8) model by integrating Generative Adversarial Networks (GANs) for data augmentation (DA) and EfficientNet-B3 for advanced feature extraction. This novel approach is evaluated using two benchmark datasets: the Animal Images Detection (AID) dataset and the Wildlife Animals Image Dataset (WAID). Through three comprehensive experimental analyses—DA, ablation, and cross-dataset validation—the proposed model is rigorously tested to demonstrate its effectiveness in urban wildlife monitoring. The results highlight the potential of the proposed model to significantly improve detection accuracy and contribute to sustainable urban wildlife conservation efforts, aligning with United Nations Sustainable Development Goals (UN SDGs)