Detecting pine wilt disease at the pixel level from high spatial and spectral resolution UAV-borne imagery in complex forest landscapes using deep one-class classification

Abstract

Pine wilt disease (PWD) poses a serious threat to the worldwide pine forest resources. Unmanned aerial vehicle (UAV) remote sensing has been widely used for PWD control, due to its flexibility and efficiency. Although pixel-level detection can obtain fine detection boundaries, there have been few related works in complex scenes because of the difficulty of setting a preset category system and the poor generalization. A preset category system establishes which categories are to be labeled, and is necessary for traditional pixel-level detection. However, the poor generalization leads to an obvious accuracy drop when detecting PWD in new scenes. In the proposed approach, to address the first issue, one-class classification (OCC) is introduced to detect diseased pixels, focusing only on the category of diseased pine trees. However, the numerous objects but low PWD pixel proportion makes the model optimization unbalanced, for which balanced unbiased detection risk estimation is proposed. To address the second issue, a novel model consisting of three-dimensional (3D) convolutional layers and transformer blocks is proposed to extract more robust features. A novel PWD detection framework based on deep OCC is finally proposed to achieve fine pixel-level PWD detection results. PWD detection experiments were conducted on eight UAV H2 (high spatial and spectral resolution) image strips. In total, 300 PWD samples from strip 1 (accounting for roughly 0.009 % of the total pixels) and 400 unlabeled pixels formed the training set. The test experiments were conducted in the remaining seven strips to validate the model generalization. Satisfactory quantitative results (F1-score greater than 0.9) were obtained for all the test strips. The results indicate that the proposed method has a powerful ability to detect PWD in pine trees, even when the PWD proportion is low, and shows better model generalization than the traditional pixel-level detection methods.