A lightweight segmentation model toward timely processing for identification of pine wood nematode affected trees with UAV

Pine wood nematode (PWN), a major international quarantined forest pest, has resulted in significant loss to the pine forest resources, posing a serious threat to global forest ecosystems. Quick and accurate identification of trees infected by PWN can lead to earlier intervention in their spread, thereby significantly reducing losses. However, there is a scarcity of algorithm that are both swift and precise. To achieve more rapid and precise segmentation of trees affected by PWN, we proposed a novel lightweight model termed Refined and Deformable Carafe Attention Net (RCANet). The RCANet excels in both accuracy and real-time performance. It has achieved segmentation accuracy that surpasses mainstream segmentation networks, including DeepLabv3 + , Segformer, PSPNet, HrNet, and UNet. The number of parameters in RCANet is only 5.373 million, the segmentation speed reached 83.14 fps. Compared to the baseline model UNet, the IoU of the affected trees class is improved by 5.6%, and the segmentation speed is accelerated by about 90%. A straightforward yet highly effective lightweight structure was proposed, termed Refined VGG. Additionally, we validate the efficacy of several network modules for this task. RCANet addressed the challenges of low accuracy and inadequate real-time capabilities in the identification of PWN-affected pine trees within intricate forest landscapes. which is expected to be deployed on UAVs in the future for real-time recognition to accelerate the identification and localization of affected trees. This work could facilitate the management of PWN.