Densenet development with squeeze-and-excitation block for tomato plant disease classification

Abstract

This study focuses on tomato leaf disease classification using an optimized deep learning architecture. This study proposes an improved architecture called DenseNet-SEGR, which integrates a novel Squeeze-and-Excitation (SE) block with a customized growth rate of 48 to improve feature selection and classification accuracy. Unlike standard methods, this model replaces Global Average Pooling (GAP) with an integral-based squeeze method, thus enabling a more continuous and accurate feature representation. The use of SE blocks dynamically recalibrates the importance of features such as texture, color, and tissue patterns, thereby increasing sensitivity to disease symptoms. The model was trained using the PlantVillage dataset, which includes 12,246 images spanning 10 tomato leaf disease categories, such as bacterial spot, early blight, late blight, mosaic virus, and healthy leaves. Various augmentation techniques, including rotation, scaling, and contrast adjustment, were employed to strengthen generalization and improve robustness against environmental variations. Furthermore, batch normalization and adaptive learning rate scheduling were integrated to enhance model stability and prevent overfitting. As a result, the DenseNet-SEGR architecture is able to achieve a classification accuracy of 98.22 %, outperforming DenseNet-121, DenseNet-201, and MobileNetV2. This result is explained by the integration of adaptive attention mechanisms, sophisticated data augmentation strategies, and optimized architecture. The results can be effectively applied in real-world precision agriculture, especially in edge-based or mobile disease detection systems for early intervention and crop protection

Plant phenotyping relevance

トマト葉の病徴を画像から分類する深層学習アーキテクチャを開発・比較しており、植物の病害状態推定が中心的な方法論的貢献である。

Code and data availability

The paper's tomato leaf disease classification uses a publicly available PlantVillage-derived Kaggle image dataset, explicitly linked by the authors. No author code or trained model is publicly deposited.