ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

Vision Transformer-Based Fusion of Hyperspectral Reflectance and Fluorescence Imaging for Citrus Leaf Disease Classification

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2600157.(doi:10.13031/aim.202600157)
Authors:   Zafar Iqbal, Thomas F. Burks, Sriram Thota, Satya A.C. Obellaneni, Quentin Frederick, Jianwei Qin, Moon Kim, Megan Dewdney
Keywords:   Citrus, HLB, leaf disease, Hyperspectral Imaging, Fluorescence imaging, Vision Transformer, HybridSN

Abstract. Global citrus production is significantly affected by leaf diseases that reduce fruit quality and yield. Accurate and rapid classification of symptomatic leaves is essential for effective disease monitoring and targeted management. In this study, a hybrid Vision Transformer (ViT)-based framework was developed for citrus leaf disease classification using hyperspectral reflectance and fluorescence imaging. Seven classes were considered, including asymptomatic, Huanglongbing (HLB), canker, melanose, scab, greasy spot, and zinc deficiency. Principal component analysis (PCA) was used for spectral dimensionality reduction, and selected reflectance and fluorescence bands were evaluated individually and in fused combinations. Modality-specific HybridSN feature extraction branches were integrated with a ViT classifier to learn complementary spatial-spectral representations.

The performance of reflectance-only, fluorescence-only, and fused reflectance-fluorescence datasets was evaluated using PCA-3, PCA-5, PCA-10, and PCA-15 band selections. Among unimodal approaches, fluorescence consistently outperformed reflectance, achieving accuracies of 0.754, 0.789, 0.892, and 0.809 compared to 0.650, 0.769, 0.791, and 0.709 for reflectance, respectively. However, both modalities showed inter-class confusion, particularly among spectrally and physiologically similar diseases such as melanose, greasy spot, scab, and HLB. In contrast, the fused RF+FL datasets achieved substantially higher performance, with accuracies of 0.997, 0.997, 0.994, and 1 across PCA-3 to PCA-15. Using the fused PCA-3 configuration, near-perfect precision, recall, and F1-scores were obtained for all classes, with only minor confusion (2%) between melanose and scab.

These results demonstrated that combining hyperspectral reflectance and fluorescence imaging provided complementary structural, biochemical, and physiological information, significantly improving disease discrimination within the hybrid ViT framework. Importantly, near-perfect performance was achieved using only three spectral bands per modality, indicating strong potential for efficient and low-cost multispectral systems for field-deployable citrus disease detection. Although fluorescence imaging remains challenging in uncontrolled field environments due to illumination constraints but may still have potential for customized portable imaging devices.

(Download PDF)    (Export to EndNotes)