Supervised Hyperspectral Band Selection Using Texture Features for Classification of Citrus Leaf Diseases with YOLOv8

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Pathology

Background:

• Citrus greening (HLB) and citrus canker cause significant economic losses in Florida citrus production.
• Early and accurate disease detection is crucial for effective management and prevention of disease spread.

Purpose of the Study:

• To evaluate the efficacy of hyperspectral imagery (HSI) and YOLOv8 for automated detection and differentiation of citrus diseases.
• To compare the influence of network size versus HSI band selection on classification performance.

Main Methods:

• Collected citrus leaves with various symptoms (HLB, canker, scab, melanose, greasy spot, zinc deficiency) and healthy controls.
• Acquired HSI data using a line-scan camera.
• Trained a YOLOv8 model on a curated multispectral dataset derived from HSI bands selected via a variance-based method.

Main Results:

• The YOLOv8 'small' network achieved a high overall weighted F1 score of 0.8959.
• Specific F1 scores for HLB and citrus canker were 0.788 and 0.941, respectively.
• Network size demonstrated a more significant impact on performance than the selection of HSI bands.

Conclusions:

• YOLOv8, utilizing intensity differences, shows promise for automated citrus disease detection.
• The model's reliance on intensity suggests robustness to variations in specific wavelength selection compared to traditional methods.