A Hybrid Deep Learning Framework for Arecanut Disease Classification Using Efficientnet-B0 And Swin Transformer

Abstract

Arecanut is an important plantation crop in India, and diseases affecting the crop reduce yield and quality. Timely identification of these diseases can help farmers take appropriate management measures. This study presents a hybrid deep learning model for the classification of five arecanut conditions: Healthy, Rot (Mahali), Split Rot, Bud Rot, and Inflorescence Dieback. The model combines EfficientNet-B0 and the Swin Transformer for image classification. The dataset consists of 3030 field images collected from arecanut plantations in Karnataka, India. Before training, the images were preprocessed using Gaussian denoising, Contrast Limited Adaptive Histogram Equalization (CLAHE), normalization, and data augmentation. The proposed model was compared with standalone EfficientNet-B0 and Swin Transformer models using the same training and testing protocol. The model achieved an overall accuracy of 94.78%, with a macro-average precision of 94.81%, recall of 94.78%, F1-score of 94.79%, and a macro-average AUC of 0.9951. The confusion matrices and ROC curves showed fewer classification errors than the individual baseline models, particularly for the Rot and Split Rot classes. The results show that combining EfficientNet-B0 and the Swin Transformer improves classification performance for arecanut disease identification.