A simple diagnostic method for citrus greening disease based on deep learning

Citrus greening disease, also known as Huanglongbing (HLB), is one of the most destructive citrus diseases, leading to branch dieback and plant death. To enhance the efficacy of this management while keeping growers’ citriculture sustainable, infected trees should be detected as early as possible. The HLB-infection can be detected precisely by PCR testing, which imposes much cost and takes several days to complete. Another problem with this method is that the precision of the pathogen detection may be dependent on probability. Leaves are collected from tested trees for chemical analyses, which presumably require the presence of the pathogen in the leaves. However, symptomatic leaves do not necessarily carry the pathogen. Recently, analyses with image data have been used to evaluate plant physiological conditions or detect causative organisms that bring specific disorders to the plant. We thus attempted to determine the HLB-infection with an artificial intelligence identification model. In this paper, we developed a brief assessment method by using transfer learning from the proposed Faster Region Convolutional Neural Network (Faster R-CNN) architecture. The architecture was trained with in-field images collected from two citrus orchards in Thailand and Vietnam which were severely infected by HLB. We tried three methods of annotating leaves as well as trees and compared the results on the test set between VGG16 and Resnet101 pre-trained models. The experiments showed that annotating HLB-infected leaves alone was less effective than adding healthy and other diseased leaves and the highest score of HLB reached 86.38% of average precision (AP) with VGG16 and 93.82% of F2-score with Resnet101. And annotating the entire canopy of the trees for training performed better on models with deeper network layers like Resnet101.