Edge-Ready Lightweight CNN Architectures for Tomato Leaf Disease Detection Using Transfer Learning

Abstract

Tomato ( Solanum lycopersicum ) is a globally important horticultural crop whose productivity is severely constrained by foliar diseases caused by fungal, bacterial, and viral pathogens. Early, accurate detection is essential for minimising yield losses and supporting precision agriculture, yet traditional diagnosis remains time-consuming, subjective, and heavily dependent on expert knowledge. With the growth of deep learning techniques, transfer learning based convolutional neural networks (CNNs) have emerged as one of the powerful tools for automation of plant disease classifications. But the main problem is that comparative analyses of multiple architectures trained under ideal conditions remain limited. This study evaluates the performances of five widely used CNN models, Inception V3, EfficientNet-B0, ResNet50, VGG16 and AlexNet. These models were fine-tuned using a curated PlantVillage tomato leaf dataset consolidated into four major classes named as: Fungal, Bacterial, Viral and Healthy. Standard preprocessing techniques, augmentation and hyperparameter settings were applied across all networks to ensure fair comparison. Experimental results have shown that Inception V3 achieved the highest accuracy (97%), followed by ResNet (95%) and EfficientNet-B0 (91%), while VGG16 and AlexNet showed low performance due to limited depth and representation capacities. Analysis of the confusion matrix indicated a consistent distinction between healthy leaves, while the primary cause of misclassification was the visual overlap between fungal and bacterial lesions. These results suggest that Inception V3 is a strong candidate for practical use in automated disease monitoring systems. Subsequent research should focus on validation in real-world settings, interpretable AI techniques, and lightweight architectures that are suitable for mobile and edge-based smart farming solutions.

Plant phenotyping relevance

トマト葉の病徴を画像から分類するCNN手法を複数モデルで比較評価しており、植物の病害状態推定と手法検証が研究の中心です。

Code and data availability

The supplied block only references the article's own hosted manuscript and the public PlantVillage dataset used as input; no paper-specific public dataset, code, model, or supplement with an authors' URL is identified.

No evidence-backed public reproduction asset is currently recorded.