A Hybrid YOLO Backbone–Vision Transformer Framework for Multiclass Mango Leaf Disease Classification

Abstract

It is crucial to have the correct diagnosis and timely detection of mango leaf diseases to minimize crop losses and ensure precision agriculture. Handcrafted features or convolutional neural networks (CNNs) have been used in traditional disease classification methods to capture local disease structure, but these traditional methods may not be able to capture the long-range spatial relationships. To tackle this problem, this research put forward a hybrid YOLO Backbone Vision Transformer (YOLO–ViT) for automatic multiclass mango leaf disease classification. The proposed architecture is a fusion of two different architectures: hierarchical feature extraction from a YOLO-based backbone and global contextual representation capability of a Vision Transformer. The framework is used to assess the extent of the following five mango leaf categories: Anthracnose, Golmichi, Healthy, Powdery Mildew and Turning Brown. The confusion matrix shows that 69 out of 75 test samples were correctly classified, resulting in an overall accuracy of 92.0%. The class-wise results suggest that there is complete recognition of Powdery Mildew and Turning Brown samples in the set of tested samples, and major class-wise misclassification is observed between the classes of Anthracnose and Golmichi. The training and validation loss curves show fast initial convergence and relatively stable training behaviour, but there is some variation in the validation loss which requires further generalization analysis. The results reveal the capability of using hierarchical representation from YOLO along with global attention from transformer for mango leaf disease classification and can serve as a foundation for the development of efficient computer vision based disease monitoring systems for precision agriculture.