Title:
Hybrid Vision Graph Network Learning for Accurate Hyperspectral Crop Disease Detection

Authors:
Mandol Md Sharif, Jiangsheng Gui*, Md Morshed Alam, Azad Mohammad Tanveer Anjum, China

Abstract:

Crop disease identification using hyperspectral imaging (HSI) remains a critical yet challenging task in precision agriculture due to limited labeled data, imbalance in disease classes, and complex spectral variations. To address these challenges, we propose Hybrid Vision Graph Network (HVGNet), a compact hybrid model that integrates a Vision Transformer (ViT), Graph Neural Network (GNN), and K‑Nearest Neighbors (KNN). The framework begins with ViT extracting global spectral-spatial representations from hyperspectral data a method inspired by successes in crop disease detection using transformer architectures. These embeddings form the basis for a KNN graph, enabling GNN to effectively model inter sample relations and enhance class separability in spectral space. Concurrently, KNN sharpens local decision boundaries, providing robustness for ambiguous cases on carefully normalized embeddings. By fusing outputs from all three components, HVGNet achieves a high accuracy of 98.2% on benchmark hyperspectral datasets, indicating its robustness to noise, imbalance, and spectral complexity. The integrated architecture offers a scalable, interpretable, and high-performing solution for in-field disease detection using hyperspectral data.

DOI: http://dx.doi.org/10.51505/ijaemr.2026.1110