Citrus Huanglongbing Detection Based on Multi-Modal Feature Fusion Learning

Area of Science:

• Agricultural Science
• Plant Pathology
• Computer Vision

Background:

• Citrus Huanglongbing (HLB), or citrus greening disease, is a global threat to citrus production with no effective treatment.
• HLB symptoms mimic nutritional deficiencies or other diseases, complicating accurate diagnosis.
• Single-source detection methods using RGB images or hyperspectral data show limited performance.

Purpose of the Study:

• To develop a robust method for recognizing HLB and differentiating it from suspected HLB, Zinc deficiency, and healthy citrus leaves.
• To enhance citrus disease detection by integrating multiple data sources.

Main Methods:

• A multi-modal feature fusion network combining RGB image analysis and hyperspectral data processing was proposed.
• Key innovations include a dimension-reduction scheme for hyperspectral data using soft attention, a bilinear fusion method for feature integration, and auxiliary classifiers.
• The network was trained and tested on a dataset of 1,325 images and 1,325 hyperspectral data samples.

Main Results:

• The multi-modal network achieved a highest accuracy of 97.89% in classifying citrus leaf types.
• Single-modal networks performed less effectively, with RGB image analysis reaching 87.98% and hyperspectral data analysis reaching 89%.
• The proposed method demonstrated superior performance over single-modal classifiers.

Conclusions:

• Multi-source information fusion using a multi-modal network offers a superior approach for detecting citrus HLB and nutritional deficiencies.
• The developed network provides an effective tool for accurate citrus disease diagnosis.