Detection of citrus greening (HLB) using ground based hyper-spectral imaging and spectroscopy

Citrus greening, also known as Huanglongbing or HLB, is a major threat to the U.S. citrus industry. Currently, scouting and visual inspection are used for screening infected trees. However, this is a time-consuming and expensive method for HLB disease detection. Moreover, as it is subjective, the current method may involve high detection error rates. The objective of this research was to evaluate the optical sensors for the detection of HLB, other diseases and nutrient deficiencies in citrus. This dissertation describes the status of citrus in Florida, current HLB status in FL, various advanced techniques for plant disease detection. It further reviews various diseases and nutrient deficiencies in citrus that may be confused with HLB. Initially, the spectral characteristics of healthy and HLB infected tree canopies were investigated. A FieldSpec spectroradiometer (350-2500 nm) was used to detect HLB-infected trees. Discriminability, spectral derivative analysis and spectral ratio analysis were used to distinguish HLB. It was found that the spectral bands of green to red and near infrared have the ability to discriminate HLB-infected trees from healthy trees. These wavelength regions include green peak wavelengths at around 530-564 nm, 710-715 nm (red edge), and near infrared wavelengths of 1041 nm and 2014 nm. In the next step partial least squares (PLS) and discriminant statistical analyses were used to identify and discriminate spectral characteristics of HLB infections in citrus trees. Results suggest that both techniques have the potential to discriminate HLB for different varieties of citrus. Overall, the full range of data gave more accurate results compared to a narrower range of reflectance data with both statistical techniques. However, the narrower, visible, range (400 nm to 900 nm) data produced better results with PLS modeling. In contrast, discriminant analysis produced better overall results with the full reflectance range. Machine learning techniques like k-nearest neighbors (KNN), logistic regression, and support vector machines (SVM) were applied for classifying the HLB data. Analysis showed that with one spectral measurement, none of the classification methods was successful in discriminating healthy from infected trees, because of the large variability in the spectral measurements. When five spectra from the same tree were used for classification, SVM and weighted KNN methods classified spectra with 3.0 and 6.5 percent error, respectively. The results from this study indicated that the canopy visible and near infrared (VIS-NIR) spectral reflectance can be used for detecting HLB infected citrus trees. However, high classification accuracy (> 90%) requires multiple measurements from a single tree. Since ASD and SVC spectroradiomters are very expensive and difficult for growers to use in field data collection, a rugged, low-cost, multi-band active optic sensor was used to identify the HLB infected trees from the healthy trees. The sensor consisted of four bands: two visible bands at 570 nm and 670 nm, and two NIR bands at 870 nm and 970 nm. Extensive field measurements were conducted using this sensor. Analysis of the data showed that due to the large variability in the data, it was not possible to discriminate healthy and infected trees based only on a single measurement from a tree. Using multiple measurements from a tree, however, it was possible to achieve high classification accuracy. With five measurements from a tree, classification methods such as k-nearest neighbors, support vector machines, and decision trees achieved classification errors of less than 5 percent. The results demonstrated the potential of a multi-band active optic sensor for detecting HLB-infected citrus trees under field conditions. This research further investigated the application of hyperspectral camera for HLB detection. Hyperspectral images of HLB infected trees and healthy trees were collected with a Specim hyperspectral camera (Autovision Inc., Los Angeles, CA, USA) having a spectral range from 306.5 nm to 1067.1 nm with 2.7 nm spectral resolution. These images were processed in ENVI 4.5 (ITT Visual Information Solutions, Boulder, Colorado). Various vegetation indices were estimated. ANOVA was used to compare the mean vegetation indices of healthy and HLB trees. Results showed that hyperspectral imaging have a potential to discriminate HLB from healthy samples.