EffResViT-SE FusionNet: A Hybrid Deep Learning Framework for Accurate Classification of Coffee Leaf Diseases.

Abstract

Coffee is a vital agricultural commodity that sustains millions of farmers worldwide, yet its cultivation is increasingly threatened by devastating leaf diseases such as Leaf Rust, Phoma, Cercospora, and Leaf Miner. These diseases reduce photosynthetic efficiency, cause defoliation, and ultimately lower crop yield and quality. Traditional diagnostic methods, including visual inspection and laboratory-based tests such as PCR and ELISA, are often time-consuming, costly, and require expert intervention, making them impractical for large-scale use. To address these challenges, we propose EffResViT-SE FusionNet, a novel hybrid deep learning framework that integrates EfficientNetB3 and ResNet50 enhanced with Squeeze-and-Excitation (SE) blocks for adaptive local feature recalibration, along with a Vision Transformer (ViT) for modeling global contextual dependencies. This fusion design effectively combines CNN-based local feature extraction with transformer-based long-range attention in a unified architecture. The model was trained on a large-scale dataset comprising 58,555 coffee leaf images distributed across five classes: Healthy (18,984), Miner (16,983), Leaf Rust (8336), Cercospora (7681), and Phoma (6571). The dataset was split into 70%, 15%, and 15% testing. Key hyperparameters included the Adam optimizer, a learning rate of 0.001, a batch size of 32, and 80 training epochs, ensuring stable convergence. Experimental results demonstrate the superior capability of the proposed model, achieving an overall classification accuracy of 99%, with precision, recall, and F1-scores all ranging between 98% and 99% across all classes. Comparative analysis confirmed notable improvements over baseline models: ResNet50 (94% accuracy), EfficientNetB3 (95% accuracy), and standalone ViT (97% accuracy). Furthermore, ablation studies validated the critical role of SE blocks and feature fusion with the transformer in achieving optimal performance. These outcomes highlight EffResViT-SE FusionNet as a powerful, precise, and scalable solution for early detection and classification of coffee leaf diseases, supporting timely interventions and promoting sustainable agriculture.

Plant phenotyping relevance

コーヒー葉画像から病害状態を分類する深層学習フレームワークの開発・比較検証が研究の中心であり、植物病害表現型の取得・推定に該当する。

Code and data availability

The paper's plant-phenotyping input is a public Kaggle coffee leaf image dataset (58,555 images across five classes) explicitly linked in the Data Availability Statement. No author analysis code or trained model checkpoints are disclosed.