A Hierarchical Transfer Learning Framework for Ginger Leaf Disease Detection and Severity Assessment Using Swin Transformer

Abstract

Ginger leaf diseases can reduce crop growth and rhizome yield when they are not identified and managed at an early stage. This study presents a hierarchical transfer learning framework based on the Swin Transformer for ginger leaf disease detection and severity assessment. The framework analyzes a leaf image in three stages. First, it separates healthy leaves from diseased leaves. Diseased leaves are then classified into five disease classes: Chlorotic Fleck, Bacterial Wilt, Fusarium Yellow, Leaf Blight, and Pyricularia Leaf Spot. In the final stage, the identified disease is assigned to one of three severity levels: Early, Middle, or Severe. A dataset of 10,500 field images has been used, including 1,750 healthy and 8,750 diseased images. The images were collected under natural daylight and included changes in illumination, background, leaf orientation, and symptom severity. The Swin Transformer models were initialized with ImageNet pretrained weights and fine-tuned for the three classification tasks. The healthy diseased classification achieved 99.43% accuracy with an AUC of 0.9984. Disease classification achieved 97.72% accuracy and a macro AUC of 0.9790, while severity classification achieved 92.40% accuracy with a macro AUC of 0.9605. The results show that the hierarchical approach can separate the identification process into manageable stages and provide both disease type and severity information from leaf images. The model can be used as a basis for image based ginger disease screening under field conditions.