A Simple Diagnostic Method for Citrus Greening Disease With Deep Learning

Abstract

Citrus greening disease (CG) is the most destructive disease of citrus, leading to branch dieback and plant death. Currently, there is no cure for CG, the early detection and removal of infected trees is important to prevent the spread of the disease. In recent years, there have been growing expectations for CG detection with digital images, especially deep learning techniques applied to digitized herbarium specimen image data. However, this approach faces challenges in practical applicability and detection efficiency. In this paper, we proposed a simple diagnostic method for CG using transfer learning with the Faster RCNN architecture. We collected in-field images from a citrus orchard in Thailand where CG has caused significant damage. We compared the performance of two annotation methods with the in-field leaf dataset and discussed their effects on pre-trained VGG and Resnet models. Five-fold cross-validation was utilized for model training and evaluation, with average precision (AP) used as the performance metric. The results showed that the Resnet models performed better than the VGG models, with the Resnet152 model scoring the highest in this task. The annotation method which included annotations of healthy and other disease leaves achieved an AP of 84.13% lower than another one but indicated better performance in practical applications with more robustness. Additionally, we developed a web application that performs real-time diagnosis using our trained models and verified the effectiveness of our system.