Early detection of pine wilt disease using deep learning algorithms and UAV-based multispectral imagery

[摘要] Pine wilt disease (PWD) is a global devastating threat to forest ecosystems. Therefore, a feasible and effective approach to precisely monitor PWD infection is indispensable, especially at the early stages. However, a precise definition of early stage and a rapid and high-efficiency method to detect PWD infection have not been well established. In this study, we systematically divided the PWD infection into green, early, middle, and late stages based on the needle color, the resin secretion, and whether the pine wood nematode (PWN) was carried. Simultaneously, an unmanned aerial vehicle (UAV) equipped with multispectral cameras was used to obtain images. Two target detection algorithms (Faster R-CNN and YOLOv4) and two traditional machine learning algorithms based on feature extraction (random forest and support vector machine) were employed to realize the recognition of infected pine trees. Moreover, we took into consideration of the influence of green broad-leaved trees on the identification of pine trees at the early stage of PWD infection. We obtained the following results: (1) the accuracy of Faster R-CNN (60.98-66.7%) was higher than that of YOLOv4 (57.07-63.55%), but YOLOv4 outperformed in terms of model size, processing speed, training time, and testing time; (2) although the traditional machine learning models had higher accuracy (73.28-79.64%), they were not able to directly identify the object from the images; (3) the accuracy of early detection of PWD infection showed an increase of 3.72-4.29%, from 42.36-44.59% to 46.08-48.88%, when broad-leaved trees were considered. In this study, the combination of UAV-based multispectral images and target detection algorithms allowed us to monitor the occurrence of PWD and obtain the distribution of infected trees at an early stage, which can provide technical support for the prevention and control of PWD.

[发布日期] 2021-10-01

[关键词] Pine wilt disease;Early detection;Multispectral imagery;Deep learning;Faster R-CNN;YOLOv4