Supervised hyperspectral band selection using texture features for classification of citrus leaf diseases with YOLOv8

Authors: FREDERICK, QUENTIN - University Of Florida; BURKS, THOMAS - University Of Florida; WATSON, JONATHAN - University Of Florida; YADAV, PAPPU - University Of Florida; Qin, Jianwei; Kim, Moon; DEWDNEY, MEGAN - South Dakota State University

Submitted to: Sensors
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 2/4/2025
Publication Date: 2/9/2025
Citation: Frederick, Q., Burks, T., Watson, J., Yadav, P., Qin, J., Kim, M.S., Dewdney, M. 2025. Supervised hyperspectral band selection using texture features for classification of citrus leaf diseases with YOLOv8. Sensors. 25(4). Article e1034. https://doi.org/10.3390/s25041034.
DOI: https://doi.org/10.3390/s25041034

Interpretive Summary: Citrus diseases cause smaller fruits, blemishes, premature fruit drop and eventual tree death, which is a major threat for the citrus industry. Since the disease symptoms generally first appear on the citrus leaves, detection and identification of the diseases via leaf inspection can permit more effective mitigation tactics and more intelligent management of the groves. This research developed an AI-based imaging detection method to identify citrus leaf diseases. A portable benchtop hyperspectral imaging system was used to collect reflectance images from both sides of Valencia orange leaves with diseases of Huanglongbing (HLB), canker, greasy spot, melanose, scab, zinc deficiency, and a control set. Vision-based object detection machine learning models (called YOLOv8) were trained using different subsets of the image data. The best classification accuracy was achieved at 89% to differentiate the seven leaf conditions. The hyperspectral imaging and the machine learning techniques would assist the citrus industry and regulatory agencies (e.g., FDA and USDA APHIS) in enforcing standards for the quality and safety of citrus-related food and beverage products.

Technical Abstract: Citrus greening disease (HLB) and citrus canker cause financial losses in Florida citrus groves via smaller fruits, blemishes, premature fruit drop and/or eventual tree death. Often, symptoms of these resemble those of other defects or infections. Early detection of HLB and canker via in-grove automated leaf inspection can enable more effective management of groves. This study tests methods to select spectral bands from hyperspectral reflectance imagery (HSI) for classifying these two conditions in the presence of other leaf defects. HSI reflectance images of both sides of citrus leaves with visible symptoms of HLB, canker, zinc deficiency, scab, melanose, greasy spot, and a control class were collected with a line-scan HSI camera. Bands from this imagery were selected using three methods, and using these bands, YOLOv8 was trained to classify these images. The ‘small’ network using an intensity-based band combination yielded an overall weighted F1 score of 0.8959, classifying HLB and canker with F1 scores of 0.788 and 0.941, respectively. The network size appeared to exert greater influence on performance than the HSI bands selected. These findings suggest that YOLOv8 relies more heavily on intensity differences than texture properties of citrus leaves and is less sensitive to the choice of wavelengths than are traditional machine vision classifiers.