Scopus Indexed Publications

Paper Details


Title
A Leakage-Aware Deep Learning Framework for Multi-Class Classification of Moringa Leaf Diseases Using CNN-Vision Transformer Architectures

Author
Rashedul Hasan Shohan, Ismail Hossain, Mst. Sazia Tahosin,

Email

Abstract

Medicinal plants play a critical role in healthcare systems, particularly in developing regions. Moringa oleifera is widely cultivated for its pharmacological value; however, its productivity is significantly affected by foliar diseases such as Bacterial Leaf Spot, Cercospora Leaf Spot, and Yellowing. Existing automated diagnosis approaches often suffer from methodological limitations, including data leakage during augmentation and the use of model architectures that are poorly aligned with domain-specific characteristics. This work presents a comprehensive deep learning framework for the multi-class classification of Moringa leaf diseases. The proposed approach integrates a strict leakage-free data splitting strategy that preserves biological sample integrity, an advanced HSV-based region-ofinterest extraction pipeline for background suppression, and a targeted transfer learning scheme employing architectures selected to assess their inductive suitability for Moringa pathology. Six ImageNet-pretrained backbones are fine-tuned, comparing convolutional neural networks (EfficientNet-B3, ConvNeXt, ResNet50, and MobileNetV3) with Vision Transformers (ViT and Swin). EfficientNet-B3 achieves the highest performance, attaining 91.78% accuracy, a 92.56% macro F1-score, and a 98.53% macro ROC-AUC. The results indicate that rigorous data preparation combined with architecture-aware model selection improves Moringa leaf disease classification; however, broader field validation and interpretability analysis remain necessary before practical deployment.


Keywords

Journal or Conference Name
2026 International Conference on Control, Automation and Diagnosis, ICCAD 2026

Publication Year
2026

Indexing
scopus