| Ground penetrating radar(GPR)has been widely used in nondestructive testing of tree roots as a geophysical technology.Compared with nondestructive testing technologies such as X-ray tomography,acoustic method and resistivity tomography,GPR has the advantages of easy operation,portable in the field,scanning large roots and repeated measurement quickly and economically.It has gradually become the preferred technology for root nondestructive testing.Due to the inevitable introduction of noise interference when collecting data in the field,the propagation speed of electromagnetic wave is lower than the theoretical value,and the clutter "blurs" the root target response on the radar image.Based on the migration imaging algorithm,combined with image processing,signal processing and deep learning technology,this thesis improves and optimizes the migration imaging algorithm,and propose three root detection and localization algorithms.The main research contents are as follows:(1)To precisely adjust the electromagnetic wave velocity for F-K migration,a root localization algorithm based on velocity correction is proposed in this thesis.The algorithm uses F-K migration combined with image entropy to correct the migration speed.Firstly,singular value decomposition(SVD)is used to remove the background of the collected GPR data,then the initial migration speed is set and the corresponding migration image entropy is calculated,and then the migration speed is gradually increased to obtain the fitting curve of the speed with respect to the image entropy.The smaller the entropy of the migration image,the better the focusing effect of the image point target.Select the wave velocity with the lowest entropy of the image as the input to test the accuracy of the corrected wave velocity.The experimental results show that SVD can remove the background of radar image well,and has an obvious effect on removing ringing noise in the measured data.The migration imaging results after correcting the wave velocity are more consistent with the reality.(2)To settle the problem that the clutter in GPR B-scan image often obscures the tree roots,thus reduces the accuracy of tree-root localization algorithm.This thesis proposes a tree root localization method combining robust depth autoencoder(RDAE),direct least squares(DLS)and F-K migration.Firstly,the zero-corrected B-scan image is decomposed into a low-rank component representing clutter and a sparse component representing root target echo by RDAE,and the sparse component is retained to complete clutter suppression;then DLS is used to fit the target echo.The relative permittivity of the soil is estimated by the hyperbolic curve.Finally,the migration velocity is calculated according to the relative permittivity of the soil as the input of the F-K migration for migration imaging,and the radius and depth information of the root are obtained to complete the root localization.The experimental results show that the clutter suppression effect of the RDAE method on the simulated and measured data has a higher signal-to-noise ratio and improvement factor.The root mean square relative error of the soil relative permittivity estimated by DLS is 3.84%.The radius relative error and the maximum depth relative error are 8.5% and 8.7%,respectively.(3)For the case of deep learning cannot explain the nature of B-scan data well,this thesis proposes a root detection and localization algorithm based on DNN back projection,which can not only detect the hyperbolic features in the original B-scan image,but also Curved features are interpreted as cross-sectional images of subterranean root targets.In addition,in order to compare with the traditional GPR data post-projection interpretation method,this thesis also introduces a synthetic GPR dataset that simulates the real NDT environment,and uses less GPR data to reconstruct the underground root target to compare Less computation yields better GPR imaging results.The experimental results show that BPNet takes less time on average than the traditional back projection,and can more accurately restore the cross-sectional shape of the tree root.In this thesis,the migration imaging algorithm is used to locate the tree roots,which improves the impact of wave velocity deviation and clutter interference on migration imaging,and the size and location information of the located tree roots are more accurate.The root algorithm is applied to the non-destructive testing of urban trees to provide decision support for the health and maintenance of ancient and famous trees. |
