Design of a GAN-Enhanced Curriculum Learning Approach for Efficient Pomegranate Growth Stage Detection Using YOLOv11m in Precision Agriculture

Tanvi Bhatnagar¹, Dr. Navneet Kumar Agrawal²

¹ M.Tech Scholar, Department of Electronics and Communication Engineering, College Of Technology and Engineering, Maharna Pratap University of Agriculture and Technology, Udaipur, Rajasthan

² Head of the Department, Department of Electronics and Communication Engineering, College Of Technology and Engineering, Maharna Pratap University of Agriculture and Technology, Udaipur, Rajasthan

ABSTRACT

Accurate detection and classification of pomegranate (Punica granatum) growth stages is critical for automated orchard management and yield estimation. Monitoring these stages allows farmers to optimize irrigation, manage pests, and predict harvest times. However, agricultural environments present severe challenges, including occlusion, varying lighting, and extreme scale variations between micro-features (buds) and mature fruits. This paper introduces a novel pipeline combining Generative Adversarial Network (GAN) augmentation with a Two-Phase Curriculum Learning approach using the highly optimized YOLO11m architecture. To combat cloud infrastructure volatility, a fault-tolerant auto-resume pipeline was engineered across Kaggle and Google Colab environments. By synthesizing rare micro-features and utilizing a progressive resolution-scaling training loop, our model achieved an exceptional Mean Average Precision (mAP50) of 97.6%, with a Precision of 95.8% and a Recall of 92.8%. This research establishes a new state-of-the-art benchmark, significantly outperforming traditional baselines such as YOLO-Punica, and provides a highly scalable framework for real-time robotic phenotyping.

INDEX TERMS Curriculum learning, deep learning, GAN augmentation, precision agriculture, multi-source datasets, YOLO11m, cloud computing.