YOLO-DRIVE INTERNET OF THINGS BASED COMPUTER-AIDED DIAGNOSIS SYSTEM FOR OLIVE LEAF DISEASES RECOGNITION

Abstract

Addressing the existing gap in the practical deployment of deep learning models for agricultural disease diagnostics, this paper presents OLIVE-CAD, an Internet of Things (IoT) enabled Computer-Aided Diagnostics system specifically engineered for the real-time recognition of olive leaf diseases. The system integrates a high-resolution camera module for image acquisition, an embedded microprocessor for on-site processing, and an IoT module for remote data transmission via PAHO-MQTT to a cloud platform. For the core diagnostic component, a YOLOv12 Convolutional Neural Network (CNN) model was developed and trained on a comprehensive dataset of 11,315 olive leaf images, categorized into 'Aculus', 'Scab', and 'Healthy'. This dataset underwent meticulous preparation, including rigorous bounding box annotation and manual review to ensure data quality. The novelty of this work lies in its end-to-end integration of a custom-built, on-site hardware solution with a high-performance deep learning model and a scalable IoT data pipeline. This provides a practical, field-deployable system rather than a theoretical model. The trained YOLOv12 model achieved a high performance, with a mean average precision (mAP@0.5) of 98.2% and class-specific accuracies of 97% for 'healthy', 84% for 'aculus', and 95% for 'scab'. These results translate into significant benefits for agricultural management, including real-time diagnostics for early disease detection, which enables swift intervention and reduces crop loss. Furthermore, the system's ability to transmit structured data provides farmers with data-driven insights for monitoring disease progression over time. System validation involved real-life data capture and processing from 60 printed samples (20 samples for each class: healthy, aculus, and scab) to simulate field conditions. The model successfully classified the majority of these samples with high accuracy, a finding that is consistent with the model's confusion matrix. The diagnostic results were smoothly transmitted to a cloud-based PostgreSQL database with TimescaleDB for efficient storage, and presented on a Grafana dashboard for real-time data viewing.