Hyperspectral Imaging and Machine Learning for Huanglongbing Detection on Leaf-Symptoms.

Area of Science:

• Plant Pathology
• Remote Sensing
• Agricultural Science

Background:

• Huanglongbing (HLB) poses a significant threat to global citrus production, leading to tree death due to a lack of effective treatments.
• Current HLB management relies heavily on removing infected trees to control disease spread.

Purpose of the Study:

• To develop and evaluate a non-destructive method for detecting Huanglongbing (HLB) in citrus using hyperspectral leaf reflectance.
• To identify key spectral features indicative of HLB infection.

Main Methods:

• Collected 72 hyperspectral leaf images from citrus trees in Thailand, categorizing them as symptomatic or asymptomatic for HLB.
• Applied Principal Component Analysis (PCA) to identify 16 characteristic wavelengths in the red-edge to near-infrared regions.
• Trained seven machine learning models (Random Forest, Decision Tree, SVM, KNN, Gradient Boosting, Logistic Regression, Linear Discriminant) using spectral data at the identified wavelengths.

Main Results:

• PCA identified 16 key wavelengths (715-736 nm and 930-997 nm) that effectively differentiate symptomatic from asymptomatic citrus leaves.
• Random Forest model achieved the highest F1-score (99.8%), demonstrating superior performance in distinguishing HLB-infected leaves.
• The top three models (Random Forest, Decision Tree, KNN) showed accuracy comparable to Polymerase Chain Reaction (PCR) results.

Conclusions:

• Hyperspectral leaf reflectance analysis, particularly at specific red-edge and near-infrared wavelengths, is a highly effective non-destructive method for HLB detection.
• Machine learning models, especially Random Forest, can accurately classify citrus leaves based on spectral data, providing a reliable alternative to PCR.
• This approach offers a promising tool for efficient and early HLB management, aiding in disease control and reducing economic losses in citrus orchards.