Classifying adaxial and abaxial sides of diseased citrus leaves with selected hyperspectral bands and YOLOv8

Abstract

Citrus greening disease (HLB) and citrus canker are diseases afflicting Florida citrus groves, causing financial losses through smaller fruits, blemishes, premature fruit drop and/or eventual tree death. Often, symptoms of these resemble those of other defects/infections. Early detection of HLB and canker via in-grove leaf inspection can permit more effective mitigation tactics and more intelligent management of groves. Autonomous, vision-based disease scouting in a grove offers a financial benefit to the Florida citrus industry. This study investigates the potential of hyperspectral reflectance imagery (HSI) for detecting and classifying these conditions in the presence of other, less consequential leaf defects. Both sides of leaves with visible symptoms of HLB, canker, zinc deficiency, scab, melanose, greasy spot, and a control set were collected and imaged with a line-scan hyperspectral camera. Spectral bands from this imagery were selected using two methods: an unsupervised method based on principal component analysis (PCA), and a supervised method based on linear discriminant analysis (LDA). Using the selected bands, the YOLOv8 network architecture was trained to classify each side of these leaves. LDA-selected bands from the back of the leaves yielded an overall classification accuracy of 84.23%. Leaves with HLB and zinc deficiency were classified most accurately, with F1 scores of 0.977 and 0.953, respectively. On the back side of the leaf, recall of melanose was significantly improved by using the LDA bands. These findings favor the use of supervised band selection for HSI-based in-grove disease detection.