A Proficient Multiple Mango Disease Classification of Mango Tree Using Metaheuristic-Assisted Recurrent Residual U-Net + +-Aided Segmentation with Vision Transformer-Based Residual Attention Network

Identifying plant disease plays a significant role in disease control to enhance the quality as well as quantity of crop yield. Mangoes are enriched with high nutrient values that help to maintain human health. Maintaining high-quality mango yields is essential for global food security, yet traditional manual identification of diseases remains inefficient, labor-intensive, and prone to error. Existing automated research often suffers from a limited diagnostic scope, focusing on leaves or fruits while their architecture lacks the capacity to identify subtle, early-stage symptoms across multiple plant organs. Therefore, a multiple mango disease classification strategy for detecting various types of diseases, such as anthracnose, powdery mildew, sooty mold, stem end rot, and so on, is introduced in this research study with higher classification accuracy using an advanced deep learning mechanism. The primary objective of this study is to develop a proficient, multi-organ diagnostic framework capable of accurately classifying diseases across the fruit, leaf, stem, and blossom of the mango tree. The proposed methodology utilizes Generative Adversarial Networks (GANs) for data augmentation to overcome data insufficiency, followed by an Adaptive Recurrent Residual U-Net[Formula: see text] (AR2UNet[Formula: see text] for precise disease segmentation. To maximize performance, the Enhanced Botox Optimization Algorithm (EBOA) is introduced to automate parameter tuning and a Vision Transformer-based Residual Attention Network (ViT-RAN) is employed for final classification. Performance metrics such as accuracy, false positives, and overall effectiveness were used for evaluation with ViT-RAN to demonstrate high efficiency compared to conventional models, like CNN and RNN. To demonstrate the proposed framework’s effectiveness, experimental validation was conducted against several existing architectures and optimization algorithms. The findings reveal that the EBOA-AR2UNet[Formula: see text] model achieves a peak segmentation accuracy of 97.43%, significantly surpassing conventional methods. Furthermore, the ViT-RAN classifier demonstrated superior computational efficiency and reduced the memory footprint by attaining high-precision results across the fruit, leaf, stem, and blossom categories. Statistical analysis confirmed that the EBOA provides faster convergence and more stable parameter tuning than standard optimizers for validating the model’s reliability in real-time agricultural monitoring.