What question did this study set out to answer?

The study aims to create an efficient framework for detecting and classifying tomato leaf diseases using deep learning.

May 2, 2026Open Access

Hybrid Deep Learning Models for Efficient Detection and Classification of Tomato Plant Leaf Diseases

OAOlayiwola AdesobaFederal University of Technology WAWaliu ApenaFederal University of Technology FDFolasade DahunsiFederal University of Technology

Key Points

The study aims to create an efficient framework for detecting and classifying tomato leaf diseases using deep learning.
Developed a deep learning framework utilizing transfer learning, ensemble learning, and knowledge distillation.
Trained five pre-trained convolutional neural networks on a dataset of 20,245 annotated tomato leaf images.
Evaluated model performance based on accuracy and F1-score.
Achieved a test accuracy of 89.86% and a ROC-AUC of 99.19% using an ensemble of the best models.
Validation accuracy exceeded 92% and F1-score reached 0.92 after applying knowledge distillation.
Reduced student model size from 121.81 MB to 108.39 MB while maintaining high performance.

Abstract

Tomato (Solanum lycopersicum) production in Nigeria is severely affected by diseases such as bacterial spot, early blight, and yellow leaf curl virus, which reduce yields and impose economic burdens on farmers, ultimately impacting the nation’s agricultural gross domestic product (GDP). Conventional detection methods are often slow, subjective, and error-prone, leading to delayed or ineffective interventions. This study develops a deep learning framework for automated detection and classification of nine common tomato diseases and healthy leaves. The approach integrates transfer learning, ensemble learning, and knowledge distillation to achieve both high accuracy and computational efficiency suitable for mobile deployment. A dataset of 20,245 annotated tomato leaf images was used to train and evaluate five pre-trained Convolutional Neural Networks (ResNet50, EfficientNetB0, MobileNetV2, MobileNetV3-Small, and MobileNetV3-Large). An ensemble of the four best-performing models achieved a test accuracy of 89.86% and a ROC-AUC of 99.19%. Knowledge distillation was then applied, transferring knowledge from ResNet50 to MobileNetV2. The distilled student model retained strong performance, with a validation accuracy above 92% and an F1-score of 0.92, while reducing the model size from 121.81 MB to 108.39 MB. These results demonstrate the potential of the framework to deliver accurate, lightweight, and deployable solutions for plant disease detection, supporting precision agriculture and enhancing food security in Nigeria.

Ask AI

Helpful

Bookmark

View Full Paper