Accurate Detection of Pine Wood Nematode Disease Based on Deep Learning Model and UAV Images

Detecting pine wood nematode disease (PWD) in complex environments is challenging for forest managers. To address this, we develop an improved Faster R-CNN model called FRBG Faster R-CNN. Our model combines ResNet50 network, FRN normalization layer and the activation function TLU to improve the feature expression ability of Faster R-CNN backbone network. Additionally, we construct a bi-directional path aggregation network (BiPAN) to fuse the context information and design a receptive field augmentation module (GPC) to enhance the detection ability for the small targets. The proposed FRBG Faster R-CNN is demonstrated on self-established multi-size MS_PWD dataset. In this dataset, collected UAV images come from heights of 60m, 100m, and 220m. Eight data augmentation methods are applied to simulate PWD diversity and complex lighting conditions. Our model achieves 95.6% accuracy, outperforming Faster R-CNN, YOLOv5, YOLOv7, and YOLOv8 at all tested heights.