YOLO-AUG: Enhancing Wildlife Monitoring on Far-Edge Devices With Dynamic Color-Space Augmentation and Efficient Learning

Edge and far-edge computing devices are increasingly employed in applications requiring localized, low-latency processing despite their limited computational resources. One such application is wild animal monitoring in agricultural environments, which is vital for crop protection and disease prevention. Deep Learning (DL)-based models, especially those designed for image recognition, provide highly effective tools for identifying and monitoring wildlife. However, their performance is significantly constrained by two key challenges: the limited computational resources of edge and far-edge devices, and the variability in image quality caused by environmental factors. These limitations can compromise the robustness and reliability of the models, reducing their effectiveness in real-world applications. To address these challenges, we empirically evaluate the feasibility and overhead of supervised, transfer-learning, and semi-supervised training regimes for YOLO-based (You Only Look Once) detectors in resource-constrained settings, quantifying the practical trade-offs between detection performance and training overhead. Within this evaluation, we define and assess YOLO-AUG, a lightweight training-time augmentation module integrated into the YOLO training pipeline that applies on-the-fly multi-color-space transformations (e.g., Hue/Saturation/Value (HSV) and Commission Internationale de l’Éclairage Lab (CIELab) variants) to improve robustness in wildlife monitoring, without modifying the detector architecture. We also analyze memory constraints and power consumption on representative far-edge hardware (e.g., Graphical Processing Unit (GPU)-equipped trap cameras) and provide practical deployment guidelines that quantify trade-offs between accuracy, training time, and resource usage. Our evaluation underscores the trade-offs between these methodologies to provide robust and efficient monitoring solutions for our boar, wolf and deer detection scenario, moving toward a more automated workflow that reduces human labeling effort in our target deployment setting, crucial for both agricultural management and wildlife conservation.