Multiclass Citrus Leaf Disease Inspection with Region Classification and Hyperspectral Imagery

Abstract

Highlights Pixels in citrus leaf images with symptoms of zinc deficiency and citrus greening were distinguished from control. Using 15 hyperspectral bands instead of 100 only lowered performance by one percentage point. Symptoms can be detected without the need for segmentation-labeled training data. A multispectral camera with filters can provide comparable performance instead of a hyperspectral system. ABSTRACT. Foliar diseases of citrus pose a significant threat to citrus production, impacting both yield and fruit quality. Accurate and efficient detection of diseased regions with machine vision is essential for effective disease management. A Convolutional Neural Network (CNN) model was designed and tested for pixel-level disease segmentation among three classes: asymptomatic/control, huanglongbing (HLB), and zinc deficiency. The first CNN was tested with various spectral band configurations. The model using 100 hyperspectral (HSI) bands achieved the highest accuracies, with 98% for asymptomatic regions, 88% for HLB, and 92% for zinc deficient regions. However, models with uniformly spaced 10 bands and PCA-selected 10- and 15-band combinations performed comparably. The model with PCA-selected 15 bands exhibited increased misclassification for asymptomatic samples, with 2% labeled as HLB and 2% as zinc deficient, compared to 1% using all HSI bands. These findings can contribute to the development of a cost-effective multispectral disease inspection system with selected bands. This approach could enable automated instance segmentation of disease symptoms on citrus leaves and thus assist citrus growers in early intervention. Keywords: Citrus, Citrus greening, Convolutional neural network, Disease inspection, Hyperspectral band selection, Hyperspectral imagery, Multiclass disease detection, Principal component analysis, Region classification.