Automated Mango Leaf Disease Identification with Improved Grey Wolf Optimizer in Hybrid MaxViT-Swin Transformer Model

Abstract

The identification of plant diseases plays a crucial role in sustaining agricultural productivity and minimizing economic losses. Traditional approaches, which often depend on visual assessment and the farmer’s experience, are typically inadequate for the timely recognition of infections, allowing diseases to progress and cause substantial damage. In overcoming these challenges, deep learning methods offer greater capability in solving complex classification problems than traditional machine learning algorithms. In this study, we propose a hybrid transformer-driven framework for high-precision disease detection on mango leaves. This approach combines mango leaf vein segmentation with transformer-based feature extraction. MaxViT and Swin models derive 512 and 768 features from each image, which are then combined to form a 1280-dimensional feature vector. The feature attention mechanism highlights the most informative components of the features, while the improved grey wolf optimizer reduces the increased dimensionality. 200 discriminative features were selected from the feature vector, and the decreasing features were classified using six machine learning classifiers. Experiments were performed on the MangoLeafBD dataset, which contains eight classes: seven diseases and a healthy class. The proposed MaxViT-Swin–IGWO hybrid framework achieved remarkable results, achieving 100% accuracy for the Linear Discriminant classifier and 99.98% accuracy for the Neural Network classifier. Performance analysis was accomplished using precision, recall, F1-score, dice, and ROC criteria. Furthermore, an ablation test was conducted to evaluate the impact of individual model variations on the preprocessing pipeline. The findings revealed that the proposed MaxViT Swin–IGWO hybrid framework detects mango leaf diseases with superior performance, outperforming both conventional and contemporary alternatives.