Comparing the Effectiveness of Novel Vision Transformers to Convolutional Neural Networks in Real Time Detection of Citrus Diseases With Integration on Drone Platform, Year II

Overview

The citrus industry in Florida has been devastated with multiple diseases - greening, canker, and blackspot - leading to tree death, orchard closure, and lost orange production. Traditional manual inspection methods are time-consuming, expensive, and inefficient. This project explores the use of real-time computer vision systems powered by artificial intelligence (AI) to autonomously detect citrus diseases using RGB imagery. Two object detection models were compared: YOLOv8, a Convolutional Neural Network (CNN), and RT-DETR, a Vision Transformer (ViT). Both models were trained on a custom benchmark dataset which features over 6,000 citrus images augmented by difficulty level to simulate real-world conditions such as lighting and motion blur. Five-fold cross validation, hyperparameter tuning, and performance analysis across the dataset revealed that YOLOv8 outperformed RT-DETR in accuracy and real-time speed. Grad-CAM visualizations were generated to compare model attention and interpret failure cases. YOLOv8 achieved 92.5% mAP versus RT-DETR’s 87.07%. The models' architecture were also optimized via neural architecture search and ablation study, and hyperparameters via bayesian optimization. The optimized YOLOv8 model was deployed on a Raspberry Pi 5 with a camera module, mounted on a consumer-grade drone. This lightweight system successfully performed live disease detection with geo-tagging in a field test over a citrus grove, showcasing accurate classification of diseased leaves / fruits using only RGB data. A real-time dashboard demonstrated detections while the drone flew through the farm. This work demonstrates the viability of low-cost drone platforms for disease detection and introduces a novel data framework for future research in AI implementation on drones.