Multi-Head Shifted Window-Based Swin Transformer Model with Kookaburra Optimization for Plant Leaf Disease Classification

Plant disease detection and classification are important for safeguarding crop health, optimizing agricultural practices, and ensuring food security. Nowadays, traditional deep learning (DL) methods have been utilized for disease detection and classification. However, they face significant challenges, such as overfitting and high computational complexity. To overcome these challenges, the study proposes a novel transformer-based model named the Multi-head Shifted Window-based Optimized Deep Swin Transformer for precise classification of plant leaf diseases. This model uses a Type 2 fuzzy Circular-based Neighborhood Cuckoo Search algorithm to enhance the input leaf image. Then, a Multi-head Shifted Window Interactive Dual-branch Swin Transformer U-Net model is used to identify the disease-infected areas, and a Convolutional Inception v3-based Pyramid Squeeze Attentional Kookaburra Network (CIv3-PSAKN) is used to extract features and classify the plant leaf images. The Kookaburra Optimization Algorithm can be used to adjust the weight of the CIv3-PSAKN model, helping to extract highly discriminative features and classify plant leaf images more efficiently. Using the PlantVillage dataset, the proposed model achieves a high accuracy of 99.38%. By addressing important shortcomings of current DL models in agricultural applications, this all-encompassing method seeks to improve the precision and effectiveness of plant disease identification. On the PlantVillage dataset, the proposed model achieves a classification accuracy of 99.39%, precision of 99.62%, recall of 99.49%, [Formula: see text]1-score of 97.51%, MCC of 95.13%, and Kappa score of 95.19%. Segmentation performance yields a Dice score of 0.9856, a mean IoU of 0.998, and a segmentation accuracy of 99.39%. Moreover, the proposed model demonstrates superior efficiency, with a prediction speed of 13 s and reduced error rates: RMSE [Formula: see text] 0.1193, MSE [Formula: see text] 0.0142, and MAE [Formula: see text] 0.2609. Total computation time is minimized to 135 s, outperforming conventional models like CNN, ResNet, and VGG in both accuracy and speed.