Semi-Supervised Learning Method for Pine Wilt Disease Detection in UAV Images

Objectives Pine wilt disease (PWD) is a global disease with devastating threat to forests, and causes great ecological and economic losses in China every year. It is important to monitor and map the infection status of pine trees accurately for industry sectors to control the spread of such diseases timely. Traditional PWD detection methods such as manual inspection and visual interpretation of remote sensing images are inefficient. Recently, deep learning algorithms are applied in PWD detection task, bringing significant performance gains. However, the current methods usually adopt supervised learning strategy, which requires lots of professional labeling that is time-consuming and labor-intensive.

Methods We propose a semi-supervised PWD detection method in a teacher-student mechanism. First, 16 310 images are collected in Wudu River Region in Yichang City, Hubei Province, with 697 labeled images for model training and 15 613 unlabeled images for testing. Then, YOLOv5 model is adopted as a student model to learn features of input images, and the learnt features are updated by exponential moving average algorithm to build a teacher model for generating pseudo-labels of unlabeled test images. The test images with pseudo-labels are mixed with train dataset and fed into the student model to optimize features and generate a stronger teacher model. With the corporation of student and teacher model, the network is more capable to learn robust features for detecting PWD in UAV images.

Results The extensive experiments demonstrate that the proposed semi-supervised method can effectively detect the diseased wood and achieve a promising performance with precision of 85.05% and recall of 91.36% on the test dataset, the ground truth of which is provided by third-party professionals.

Conclusions The proposed semi-supervised object detection method uses a small portion of labeled data to robustly learn features in large scale UAV images for high-quality PWD detection. The performance on test dataset shows that the proposed method could be well applied to the application of rapid dynamic monitoring of PWD epidemics in forestry field, which could benefit for the sustainable development of forest ecosystem.