Swin-HViT: A Hybrid Transformer Approach for Accurate Early-Stage Crop Disease Diagnosis

Abstract

Abstract Agriculture plays a pivotal role in global economic growth, yet it faces significant challenges from pests and crop diseases. Early detection is crucial for preventing large-scale crop losses and ensuring food security. This study introduces a hybrid transformer model, Swin-HViT, which integrates the strengths of Vision Transformer (ViT) and Swin Transformer to accurately predict crop diseases. While ViT captures global image features, Swin Transformer excels at extracting fine-grained local details. Evaluated on two benchmark datasets, Corn and PlantDoc, our model achieved accuracy of 98.81% and 81.81%, respectively, surpassing recent works. Here, we demonstrate the effectiveness of combining complementary transformer architectures to enhance disease identification in diverse agricultural settings. The code, data and the hybrid model are available at https://github.com/hema2107/Swin-HViT.

Code and data availability

The paper explicitly states that the code, data, and hybrid model are publicly available in the authors' GitHub repository, and it evaluates on two public Kaggle plant-disease image datasets (Corn/maize leaf disease and PlantDoc) that serve as the phenotyping image inputs for the study.