Citrus Huanglongbing disease identification using computer vision and machine learning.

摘要

The insect-spread bacterial infection known as Huanglongbing (HLB) or citrus greening is a very destructive citrus disease and has caused massive losses in Florida's citrus industry. No effective cure for this disease has been reported yet, and the HLB-infected tree will eventually die. Therefore, the infected tree must be detected and removed immediately to stop the spread of the disease. Early, easy, and less expensive HLB detection based on particular symptoms, such as starch accumulation in the citrus leaf, would increase the chance of preventing the disease from being spread and causing more damage. The ability of narrow-band imaging and polarizing filters in detecting starch accumulation in symptomatic citrus leaf was evaluated in this dissertation. Two custom-made image acquisition systems were developed for this purpose. In the first prototype, leaf samples were illuminated with polarized light using narrow-band high-power light emitting diodes (LED) at 400 nm and 591 nm, and the reflectance was measured by two monochrome cameras. Two polarizing filters were mounted in perpendicular directions in front of the cameras so that each camera acquired an image with reflected light in only one direction (parallel or perpendicular to the illumination polarization). Overall average accuracies of 93.1% and 89.6% in HLB detection were obtained for the 'Hamlin' and 'Valencia' varieties, respectively, using a step-by-step classification method. The second prototype was a vision sensor that included a highly sensitive monochrome camera, narrow band high power LEDs, and polarizing filters. The sensor was first tested and calibrated in a simulated field condition in a laboratory. Then, it was tested in a citrus grove. HLB detection accuracies which ranged from 95.5% to 98.5% were achieved during the laboratory and field experiments. The vision sensor images were compared with the images captured by a color camera to demonstrate the improvement achieved in this method. Also, the starch accumulation identification was studied for citrus leaves before and after being ground. The results showed an enhanced HLB identification performance using the developed vision sensor.