USING REMOTE SENSING AND GIS FOR DETECTING AND MAPPING INVASIVE WEEDS IN RIPARIAN AND WETLAND ECOSYSTEMS
Title: Spectral difference analysis and airborne imaging classification for citrus greening infected trees
Submitted to: Computers and Electronics in Agriculture
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: March 15, 2012
Publication Date: April 2, 2012
Citation: Li, X., Lee, W.S., Li, M., Ehsani, R., Mishra, A.R., Yang, C., Mangan, R.L. 2012. Spectral difference analysis and airborne imaging classification for citrus greening infected trees. Computers and Electronics in Agriculture. 83:32-46.
Interpretive Summary: Citrus greening, also known as Huanglongbing (HLB), is a serious citrus disease that affects all citrus cultivars and causes rapid deterioration of citrus trees. Early detection is very critical for the management and control of the disease. Although polymerase chain reaction (PCR) is so far the most accurate method to confirm HLB, it is labor intensive and time consuming. This study compared several image classification methods for identifying HLB-infected citrus trees from high resolution airborne multispectral and hyperspectral imagery taken from citrus orchards in Florida. Results showed that high resolution multispectral and hyperspectral imagery in conjunction with image classification techniques can identify infected trees with accuracies of 60 to over 90% in the validation data sets among the classification methods.
Citrus greening, also called Huanglongbing (HLB), became a devastating disease spread through citrus groves in Florida, since it was first found in 2005. Multispectral (MS) and hyperspectral (HS) airborne images of citrus groves in Florida were acquired to detect citrus greening infected trees in 2007 and 2010. Ground truthing including field and indoor spectral measurement, infection status along with GPS coordinates was conducted for both healthy and infected trees. Ground spectral measurements showed that healthy canopy had higher reflectance in the visible range, and lower reflectance in the near-infrared (NIR) range than HLB infected canopy. Red edge position (REP) also showed notable difference between healthy and HLB canopy. But the difference in the NIR range and REP were comparably more sensitive to the environment or the background noise. Accuracy for separating HLB and healthy samples reached more than 90% when a simple REP threshold method was implemented in the ground reflectance datasets, regardless of field or indoor measurement; but it didn’t work well with the HS images because of its low spatial resolution. Support vector machine (SVM) was able to provide a fast, easy and adoptable way to build a mask for tree canopy. High positioning error of the ground truth in the 2007 HS image led to validation accuracy of less than 50% for most of classification methods. In the 2010 image from Southern Gardens (SG) grove, with better ground truth records, higher classification accuracies (about 90% in training sets, more than 60% in validation sets for most of the methods) were achieved. Disease density maps were also generated from the classification results of each method; most of them were able to identify the severely infected areas. Simpler classification methods such as minimum distance (MinDist) and Mahalanobis distance (MahaDist) showed more stable and balanced detection accuracy between the training and validation sets in the 2010 images. Their similar infection trend with ground scouted maps showed a promising future to manage HLB disease with airborne spectral imaging.