Monitoring Method Of Pine Wood Nematode Disease Based On Multi-source Remote Sensing Data

Pine wilt disease is considered the ‘cancer’ of pine trees,with high infectivity and mortality rates,and has become a widespread epidemic threatening forest safety.Timely monitoring and detection of infected pine trees and using scientific methods to treat them is of great significance in preventing the further spread of the disease.To address the problem of the current reliance on manual ground monitoring for monitoring pine wilt disease,which has poor timeliness during general surveys,unmanned aerial vehicles(UAVs)equipped with visible light,laser radar point cloud,and ground data were used to study precise monitoring methods for pine wilt disease using multi-source data.This provides technical support for efficiently and accurately obtaining information on infected pine trees and timely treatment.(1)UAVs equipped with laser radar imaging systems and visible light cameras were used to collect visible light and laser radar data in the forest,and ground pine tree position information was obtained using handheld RTK.The obtained UAV remote sensing data was pre-processed to generate visible light orthoimages and radar point cloud images.A data set for object detection and semantic segmentation models was constructed through data augmentation and image annotation.(2)To address the problems of background interference and small target size in monitoring pine wilt disease using visible light data,SE,CA,and CBAM attention modules were introduced into the backbone network of the YOLOv5 s model to construct a pine wilt disease detection model.The m AP test results for the YOLOv5 s,YOLOv5s-SE,YOLOv5s-CA,and YOLOv5s-CBAM models were 0.797,0.801,0.808,and 0.814,respectively.The results showed that introducing attention modules into the YOLOv5 s model helps improve the accuracy of pine wilt disease detection.To address the problem of irregular pine tree crown shapes caused by tree crown occlusion in remote sensing data,the deformable convolution(DCNv2)module was introduced into the YOLOv5s-CBAM model.The improved model,YOLOv5s-CBAM-DCNv2,had an average precision mean average precision(m AP)of 0.817 for pine wilt disease detection,which was 0.3% higher than that of YOLOv5s-CBAM.Therefore,the improved model can more accurately detect pine wilt disease.(3)U-Net and R2U-Net semantic segmentation models were used to segment infected pine trees.The segmentation accuracy of the R2U-Net model for infected trees was 0.921,which was 0.008 higher than that of the U-Net model.The accuracy of segmenting diseased and dead trees was 0.919,which was 0.017 higher than that of the U-Net model.The similarity of the segmented infected trees was 0.831,which was 0.093 higher than that of the U-Net model,and the similarity of the segmented dead trees was 0.879,which was0.059 higher than that of the U-Net model.To address the problem of unclear boundaries of infected pine trees,the attention gates(AGs)module was improved by replacing the Re LU activation function in the AGs attention model with the scaled exponential linear unit(Se LU)activation function and moving the original model output to before the resampling operation.The improved AGs were introduced into the R2U-Net model to construct the ATTR2U-Net model.By introducing improved AGs into the R2U-Net model,the ATTR2U-Net model was constructed.The accuracy,recall,and similarity of ATTR2 UNet for the segmentation of pine trees with pine wilt disease were 0.960,0.856,and 0.876,respectively.The accuracy,recall,and similarity for the segmentation of dead pine trees with pine wilt disease were 0.951,0.811,and 0.901,respectively.Compared with the original model,the improved model increased the segmentation accuracy of pine trees with pine wilt disease by 3.9%,1.4%,and 4.5%,respectively,and the segmentation accuracy of dead pine trees with pine wilt disease by 3.2%,5.8%,and 2.2%,respectively.(4)The accurate location information of pine trees with pine wilt disease in the experimental area was extracted from multi-source data.The segmentation accuracy of three methods for single-tree extraction using point cloud segmentation was analyzed.The results showed that the single-tree segmentation based on normalized point cloud had the best effect,with a recall rate of 0.827,accuracy rate of 0.786,and precision of 0.806 in the high-density area;a recall rate of 0.894,precision of 0.853,and accuracy of 0.873 in the medium-density area;and a recall rate of 0.945,precision of 0.912,and accuracy of 0.929 in the low-density area.The single-tree parameters obtained by the point cloud normalization segmentation method were used to produce a tree crown Shapefile file with geographic location information.By combining the semantic segmentation model and the GDAL library to obtain the position data of 338 pine trees with pine wilt disease,9 ground targets that were misclassified as pine trees with pine wilt disease by the semantic segmentation were separated,and the location information of 329 effective pine trees with pine wilt disease caused by pine wood nematode was obtained.The obtained location information of pine trees with pine wilt disease was verified using the real geographic coordinates of the diseased pine trees obtained through ground investigation,and the results showed that all 162 pine trees with pine wilt disease investigated were located within the buffer crown obtained by the point cloud single-tree segmentation.Therefore,the point cloud single-tree segmentation algorithm combined with the semantic segmentation model can effectively remove misclassification of pine trees with pine wilt disease and improve the monitoring accuracy of pine wood nematode.