Enhanced Drone-based Plant Disease Detection with Visual Data Augmentation and Optimized Training

This paper presents an enhanced detection model utilizing data augmentation and learning rate optimization over the YOLO detection framework, applied to drone-captured imagery of cashew orchards. We focused on three stressors: insect damage, biotic stress from pathogens, and abiotic factors. To address dataset imbalance, advanced data augmentation was implemented alongside a dynamic learning rate adjustment for optimal performance. We evaluated two object detection models, YOLOv9 and YOLOv10. YOLOv9 achieved an mAP50 of 0.591 for insect detection, 0.512 for abiotic stress, and 0.512 for disease detection. YOLOv10 improved abiotic detection with an mAP50 of 0.773 but showed lower performance for insect stress (0.464) and disease (0.476). These results demonstrate the improvements achieved through our enhancements, highlighting the need for diverse datasets and better visualization techniques in precision agriculture.