Citrus Greening
Citrus greening is an economically important disease of citrus plants around the world. In Nepal, the disease was first identified in 1968, and by now the disease has spread across all the major citrus-producing hubs of Nepal. The disease is caused by Candidatus Liberibacter spp bacteria, however, it is spread from one tree to another by an insect named Asian Citrus psyllid. Citrus greening does not have a cure. Therefore, the only thing that can be done about it is to prevent it from spreading; in other words, identify the infected trees and cut them down. The complication is in identifying the infected trees. In Nepal, the only method available to confidently identify the infected trees is leaf-by-leaf polymerase chain reaction(PCR) testing. The bottleneck for PCR testing is primers which in the context of Nepal have to be imported from the US or South Korea. Primers can take months to get to Nepal consequently causing PCR testing to take up several months. To make the problem worse, in the early asymptomatic stage, Citrus greening is localized in the few leaves of the trees which further complicates sampling of the leaves for PCR testing.
The visible symptoms of Citrus greening are blotchy mottle leaves, stunted growth, premature fruit drop, reduced fruit size, corky vein, and root decline. However, the infections start from a few localized leaves. Furthermore, the infected trees can remain asymptomatic in the visible light spectrum for months to years.
Therefore, due to the multispectral drone's unique ability to capture pictures in a broader light spectrum such as Reg Edge(RE) and Near Infrared(NIR), in the last few years around the world, the usage of multispectral drones has been popular to scan citrus orchards. The utility of the technology can be extremely high especially in the context of Nepal because first, the existing method to diagnose the disease takes too long, and second, the hilly terrain in which citrus farming is done is arduous to navigate by human surveyors. We are the first team to have brought a multispectral drone to Nepal, and we are using the multispectral images to develop algorithms to rapidly detect citrus greening on orange trees.
Preliminary results of Citrus disease detection
Here we discussed the procedure used to detect citrus diseases.
Steps
1. We collected the UAV flight images (RGB)
2. We stitch the image and generate an RGB orthomosaic
3. We then manual encircle the disease and healthy citrus plant
4. We extract the pixel value (RGB ) for each healthy citrus plants
In this way, we get 986 positive pixels and 143 negative pixels
5. We trained a Random Forest classifier using these data. While training the classifier, we set 20% pixel as the test set and evaluate the results.
6. The experimental results show that the random forest classier has an accuracy of 88.15% only.
Please note that A denotes positive class and B represent negative class.
7. We suspect that the classifier is overfitted as we have used only pixels from two citrus trees (one positive and one negative). We will add more citrus tree sample and train the classifier to detect citrus disease in future.
