Edge–Cloud Collaborative Vision Framework for Multi-Class Lemon Leaf Disease Detection Using YOLOv8 and EfficientNet-B3

Abstract

The field conditions of citrus leaf disease diagnostics have not been correctly and real-time diagnosed because of complex backgrounds, variability in illumination, and overlapping foliage. The paper introduces an edge-cloud collaborative vision model of the multi-class lemon leaf disease detection system based on a real-time object detection approach and high-accuracy classification. The system suggested is implemented to deploy YOLOv8-Tiny to the edge to detect lemon leaves and extract the region-of-interest and classify the disease with a fine-tuned EfficientNet-B3 in the cloud. The two stage design eliminates background interference that is characteristic in single stage classification methods used on raw field images.
A self-made collection of lemon leaf pictures in varying climatic and lighting situations has been gathered. The diffusion-based generative models were applied to perform the generalization and reduce the imbalance between classes by generating synthetic samples to enrich the training data. The detection model was trained to be deployed on the edge, whereas the classification model was made into a TensorFlow Lite model and optimized through post-training quantization.
Experimental results indicate that the suggested framework can establish an excellent detection performance with a median Average Precision (mAP50 ) of about 95% and multi-class classification of disease with an accuracy of about 97%, and with an end-to-end inference latency of about Windows, which is low enough to support real-time agricultural monitoring. Field-level validation is an indication that the system is consistent in its performance in real-life conditions. The paper presents the edge-cloud architecture as a scalable and energy-efficient design to citrus disease surveillance in agriculture that is limited in resources.