Optimized Inception-v4 CNN Combined with YOLOv8 andXGBoost for Tomato Plant Disease Recognition and Classification

Abstract

This study proposes an integrated hybrid pipeline approach which is the combination of an optimized Inception-v4 Convolutional Neural Network (CNN), YOLOv8 (You Only Look Once version 8) and extreme gradient boosting (XG Boost) models to detect and classify Tomato plant diseases effectively and accurately. The proposed methodology is based on the fine-grained localization ability of the YOLOv8 model to accurately localize the affected area of the leaf, multi-scale deep feature extraction ability of the optimized Inceptionv4 CNN, and XG Boost to reduce dimensions and optimize features. The hybrid model combines the Inception-v4 CNN, YOLOv8, and XG Boost models with an average CPU processing time of about 200 milliseconds per image, the model performs better in terms of computational efficiency, robustness, and faster prediction. Experimental tests on Plant Village show that the model outperforms the state of the art in 10 disease categories. Specifically, the integrated hybrid system achieved high precision, recall, F1-score, mean Average Precision (m AP), and its training accuracy was more than 0.98 (up to 0.9969), and the testing accuracy was more than 0.96 (up to 0.9695). The hybrid framework shows good reliability even for diseases that are visually similar, such as bacterial spot, early blight, late blight, mosaic virus and yellow leaf curl virus. The optimization of batch size to 32, along with the tuning of the learning rate further improved the stability of training, convergence speed and the generalization of the overall model. The proposed system holds promise for real-time, scalable, and sustainable precision agriculture, aiming for early disease detection and yield protection. Future work includes incorporating Internet of Things (IoT) edge devices, in field environmental monitoring, multimodal data integration and applying Explainable Artificial Intelligence (EAI) techniques to enhance model interpretability for end-users.