Predictive Quality Management in Nepalese Attention-Enhanced Lightweight YOLOv8 Framework for Real-Time Weed Detection in Precision Agriculture

Abstract

Weed infestation continues to be a significant biological threat to global crops, leading to losses between 30% and 70%. This often requires herbicide use, which can negatively impact the environment. This paper introduces AEL-Yolov8 (Attention-Enhanced Lightweight Yolov8), a deep learning framework designed for real-time weed detection in precision agriculture. The model integrates three main innovations into the Yolov8 backbone: (1) Coordinate Attention (CA) modules that capture spatial relationships along both axes; (2) a Bidirectional Feature Pyramid Network (BiFPN) neck that enhances multi-scale feature fusion beyond the standard Panet; and (3) Complete Intersection over Union (CIoU) loss for more accurate bounding box predictions. Additionally, structured channel pruning reduces the model size by 16%, from 43 million to 18 million parameters. Tests on a publicly available sesame crop-weed dataset (300 images, 512×512 pixels) show that AEL-Yolov8 achieves 97% accuracy, 97% precision, 96% recall, 96% F1-score, and 95% map at 0.5 Iou, outperforming the baseline Yolov8 by 1.7% in map@0.5 (t=4.21, p < 0.001). The model runs in real-time at 31 FPS on NVIDIA Jetson Orin hardware, with a latency of 32 ms, and uses about 8-8.4 GB of memory and roughly 15 W of power. Extensive ablation studies, confusion matrix analysis, precision-recall curves, and tests under various lighting conditions demonstrate its robustness and adaptability. AEL-Yolov8 meets three key criteria: detection accuracy above 95% (map@0.5), a compact size under 20 million parameters, and real-time operation exceeding 30 FPS on edge devices, making it suitable for agricultural drones, autonomous weeding robots, and embedded vision systems.