Using deep learning techniques in herd analysis and management with unmanned aerial vehicles (UAVs)

With the Agriculture 4.0 revolution, the integration of unmanned aerial vehicles (UAVs) and deep learning techniques is opening the door to significant innovations in the livestock sector. This thesis aims to enable the effective and accurate monitoring of animal herds by combining UAV imagery with modern object detection algorithms and data augmentation strategies. A multi-scale dataset was created using an SJRC F22S Pro 4K UAV at various altitudes in the Kozluk district of Batman, consisting of 131 goat and 1,531 cattle images. This dataset is notably rich in terms of both small object sizes (32–64 pixels) and variable altitude conditions (0.5 m – 30 m), making it a rare contribution to the existing literature. Within the scope of the study, eight major object detection algorithms—SSD, RetinaNet, YOLOv3, DETR, Faster R-CNN, Deformable DETR, EfficientDet, and Cascade R-CNN—were compared in detail alongside various variants of the YOLOv5, YOLOv8, and the latest YOLO11 architectures. In total, over 20 model configurations were evaluated. The Cascade R-CNN model achieved the highest mAP on the cattle dataset (89.8%), while YOLOv8m demonstrated a balanced and strong performance across both datasets (Cattle: 92.3%, Goats: 70.8% mAP50). The impact of modern data augmentation techniques (CutMix, CutOut, MixUp, Mosaic) on model performance was analyzed in depth. It was observed that hybrid strategies combining Orijinal data with augmentation methods significantly improved accuracy rates. For instance, in the goat dataset, the Orijinal + Mosaic combination enhanced Cascade R-CNN's performance by 11.4%, reaching 0.751 mAP. In the cattle dataset, Orijinal + CutOut achieved the best result with 93.0% mAP50. An average loss of 22.4% mAP was observed in high-altitude images, while at lower altitudes, the decrease was approximately 5%. These findings highlight the critical importance of aligning model training with object scale and operational altitude. In conclusion, this thesis provides substantial contributions through comparative analysis of different architectures, model performance under challenging agricultural conditions, and the effectiveness of data augmentation strategies. Enhancing models like YOLOv8m and Cascade R-CNN with careful preprocessing and augmentation approaches shows that flexible, data-specific model selection has the potential to improve the accuracy, speed, and generalization capacity of UAV-based livestock monitoring systems. Key Words: Unmanned Aerial Vehicle (UAV), deep learning, herd analysis, data augmentation, Agriculture 4.0, object detection