Papaya Leaf Disease Detection Using a Hybrid Optimized Vision Transformer with Attention Mechanisms

Abstract

Early and precise identification of papaya leaf diseases plays a crucial role in enhancing crop productivity and minimizing agricultural losses. However, existing approaches face challenges in handling complex leaf textures, irregular lesion distributions, and environmental variability. To address these challenges, this study introduces a Vision Transformer (ViT)-based classification framework combined with a hybrid Falcon-Bowerbird Optimization (FBO) strategy for papaya leaf disease recognition. The proposed pipeline utilizes histogram equalization to enhance image contrast and improve feature visibility. The ViT model employs patch-based self-attention to capture local and global dependencies across leaf regions, supporting discriminative feature learning. The hybrid FBO mechanism integrates Satin Bowerbird Optimization, Falcon Optimization, and AdaBelief to optimize hyperparameters and stabilize convergence during training. Experiments were conducted on the BDPapayaLeaf dataset, which contains five categories: Anthracnose, Bacterial Spot, Curl, Ring Spot, and Healthy leaves. The proposed model achieved an accuracy of 97.83%, outperforming the evaluated machine-learning and deep-learning baselines, including CNN, AlexNet, and MobileNetV3. In addition, Grad-CAM, Grad-CAM++, and LIME were used to provide qualitative insight into image regions contributing to the model’s predictions. The results demonstrate strong classification performance under the evaluated experimental conditions and indicate the potential of the framework for future AI-assisted precision-agriculture applications, subject to further large-scale field validation.