Fusionnet: Multi-Backbone Feature Fusion Deep Neural Network for Tomato Leaf Disease Classification

Abstract

Plant diseases cause a substantial reduction in crop yield, thus affecting food security in the farming industry. The timely discovery and precise diagnosis of the diseases on tomato leaves is very important to take corrective measures. This paper proposes FusionNet, that combines several backbones for identifying diseases on the leaves of tomatoes. The proposed model combines the complementary feature learning capabilities of MobileNetV3-Small, SE-ResNet50, and ECAResNet50d through a concatenation-based fusion approach to comprehensively learn fine-grained texture details, global contextual semantics, and channel-attentive information. The proposed framework is evaluated on a curated dataset of$\mathbf{1 8, 1 6 0}$images from ten classes. To guarantee statistical significance, 5fold cross-validation with two separate runs per fold is conducted, achieving a mean validation accuracy of$99.30 \% \pm \mathbf{0. 1 9 \%}$, indicating sTable generalization. To validate the fusion strategy, t-SNE visualizations show enhanced inter-class separation as well as intra-class compactness in the fused feature space compared to the individual backbones. Cosine similarity analysis also confirms a decrease in inter-class correlation and an improvement in the discriminative structure. The experimental results show that FusionNet achieves robust, reliable, and highly discriminative performance for automated plant disease diagnosis in precision agriculture applications.

Plant phenotyping relevance

トマト葉の画像から病害状態を推定する深層学習手法を開発・評価しており、植物病害フェノタイピングが中心である。

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.