Open Access

ARTICLE

Multi-Stage Vision Transformer and Knowledge Graph Fusion for Enhanced Plant Disease Classification

Wafaa H. Alwan^1,*, Sabah M. Alturfi²

1 Computer Science, Faculty of Computer Science & Information Technology, University of Karbala, Karbala, 56001, Iraq
2 Law College, Karbala University, Karbala, 56001, Iraq

* Corresponding Author: Wafaa H. Alwan. Email:

Computer Systems Science and Engineering 2025, 49, 419-434. https://doi.org/10.32604/csse.2025.064195

Received 08 February 2025; Accepted 03 April 2025; Issue published 30 April 2025

Abstract

Plant diseases pose a significant challenge to global agricultural productivity, necessitating efficient and precise diagnostic systems for early intervention and mitigation. In this study, we propose a novel hybrid framework that integrates EfficientNet-B8, Vision Transformer (ViT), and Knowledge Graph Fusion (KGF) to enhance plant disease classification across 38 distinct disease categories. The proposed framework leverages deep learning and semantic enrichment to improve classification accuracy and interpretability. EfficientNet-B8, a convolutional neural network (CNN) with optimized depth and width scaling, captures fine-grained spatial details in high-resolution plant images, aiding in the detection of subtle disease symptoms. In parallel, ViT, a transformer-based architecture, effectively models long-range dependencies and global structural patterns within the images, ensuring robust disease pattern recognition. Furthermore, KGF incorporates domain-specific metadata, such as crop type, environmental conditions, and disease relationships, to provide contextual intelligence and improve classification accuracy. The proposed model was rigorously evaluated on a large-scale dataset containing diverse plant disease images, achieving outstanding performance with a 99.7% training accuracy and 99.3% testing accuracy. The precision and F1-score were consistently high across all disease classes, demonstrating the framework’s ability to minimize false positives and false negatives. Compared to conventional deep learning approaches, this hybrid method offers a more comprehensive and interpretable solution by integrating self-attention mechanisms and domain knowledge. Beyond its superior classification performance, this model opens avenues for optimizing metadata dependency and reducing computational complexity, making it more feasible for real-world deployment in resource-constrained agricultural settings. The proposed framework represents an advancement in precision agriculture, providing scalable, intelligent disease diagnosis that enhances crop protection and food security.

---

Keywords

---