A Hybrid Deep Learning Architecture for Ginger Disease Detection Using Leaf Images

Abstract

Ginger (Zingiber officinale) is an economically important spice and medicinal crop that suffers significant yield losses due to foliar diseases, necessitating accurate and timely automated detection systems. This study presents a hybrid deep-learning architecture for the classification of ginger leaf diseases under natural field conditions. The proposed model classifies ginger leaf images into four major disease categories: Bacterial Wilt, Fusarium Yellow, Leaf Blight, and Leaf Spot. A field-acquired dataset of 2560 images captured under diverse illumination and background conditions in India was compiled and expanded to 8950 training samples through systematic augmentation. EfficientNetB0 serves as the convolutional backbone for local feature extraction, whereas a transformer encoder captures global contextual features; the two representations are fused via concatenation prior to final classification. The proposed hybrid EfficientNetB0–Transformer model achieved a mean classification accuracy of 96.17% ± 0.31% and a macro F1-score of 0.9586 ± 0.0028 across five independent training runs with different random seeds, outperforming classical machine learning and deep learning baselines, including SVM (78.00%), BPNN (79.25%), and Vision Transformer (93.75% ± 0.48%). Notably, the proposed model surpassed the 95.2% accuracy reported in prior ginger disease detection studies while extending the classification to four disease categories simultaneously with a moderate parameter footprint of 10–12 million. Gradient-weighted Class Activation Mapping (Grad-CAM) visualisations confirm that the model consistently attends to biologically meaningful leaf regions—lesion boundaries, chlorotic patches, and vascular discolouration—providing interpretable evidence of its decision-making process. These results demonstrate that integrating convolutional local feature learning with transformer-based global contextual modelling provides an efficient, robust, and interpretable approach for automated ginger leaf disease diagnosis under real field conditions.