Efficient Wheat Disease Identification Using Hybrid Swin-SHARP Vision Model

Accurate identification of wheat diseases is an essential component for increasing crop yields and guaranteeing global food security. However, subjective opinions, errors, and laborious procedures frequently limit traditional approaches, which are based on expert knowledge. To address these challenges, we propose a Swin-Streamlined High Accuracy and Reduced Parameters (Swin-SHARP) Transformer for classifying wheat diseases. The feature extraction capability of the Swin-SHARP model is enhanced by optimizing hierarchical representations and eliminating the requirement for intricate layers and attention heads. The proposed model reduces the parameters by 82.5%, from 48.9 to 8.5 million, without compromising accuracy. A Swin-SHARP Vision Hybrid (SVH) model integrates two powerful transformer architectures, the Swin-SHARP transformer that uses shifted windows, and the Vision Transformer (ViT) for wheat disease classification. The model features a customized convolutional auto-encoder decoder that inputs 3-dimensional features. The dual convolutional layer in the encoder with a stride of 2 and kernel size 3 is activated by ReLU, that maps a 3-channel input to 16 and then 32 feature maps. This architecture reduces dimensionality by 33% while retaining crucial information for precise disease detection. Our proposed model achieved an accuracy of 98.2% on the Wheat Diseases dataset, which is also cross-validated using the YELLOW-RUST-19 benchmark by achieving 97.2% accuracy. The proposed model outperformed state-of-the-art existing models, including CNN-CGLCM-HSV-SVM, GhostNet V2, ResNet 50, MobileNet V2–VGG-16, CNN-CGL CM-HSV-SVM, and Swin Transformer by 4.8%, 1.76%, 1.2%, 13.6%, 4.8%, and 1.4% respectively, on the YELLOW-RUST-19 dataset. These findings concluded that the SVH model is a strong approach to accurately detect yellow rust disease and can be used in real time, particularly in resource-constrained agricultural settings.