Research On Tree Root Detection Algorithm Using Ground Penetrating Radar

Roots play a vital role in plant individuals.It can not only absorb water and nutrients,support the growth of plants,but also stabilize the plant body and prevent it from lodging.However,they are more challenging to observe and sample than the subaerial parts of trees.Ground Penetrating Radar(GPR)is an emerging non-destructive testing technology that has the advantages of simple operation,easy portability,and repeatable measurement compared to other non-destructive testing methods(such as X-ray tomography,nuclear magnetic resonance method,acoustic method,and electrical resistivity tomography.),so it is widely used for the detection of tree roots and trunks.However,there are difficulties in root detection,localization,and prediction due to the lack of publicly available GPR root data sets and the interference of noise and clutter that are unavoidable in real data.To address these problems,this paper combines generative adversarial networks,deep learning,signal processing and other techniques to propose three root detection algorithms,each continuing to advance based on the previous one.The main research contents are as follows:(1)To address the problem of low accuracy of root target detection due to the lack of GPR root data sets,this paper proposes a root target detection algorithm based on Generative Adversarial Networks(GAN)and the Swin Transformer.Firstly,generative adversarial networks generate highquality and sufficient data as data sets.Then,after manual labeling,the Swin Transformer extracts features from the data.Then,Feature Pyramid Network(FPN)is used to achieve multi-scale feature fusion,which improves the performance of detecting small root targets.Finally,Region Proposal Network(RPN)demarcates root targets.Simulation experiments and field experiments show that the simulated data set generated by generative adversarial networks can effectively complete the training of the root target detection network.On real data,the recall of the root target detection network based on the Swin Transformer is 0.987,and m AP is 0.871,which has better comprehensive performance than similar models.(2)To address the problem of noise and clutter affecting root localization accuracy in GPR BScan images,this paper proposes a root localization method based on Robust Principal Component Analysis(RPCA),Singular value decomposition(SVD)and Direct least square(DLS).Firstly,the BScan image is decomposed into two matrices: low-rank and sparse,where the sparse matrix is the noise extracted by RPCA.Then,SVD is used to remove background noise from the low-rank matrix.Next,the region of interest of the reflection hyperbola is obtained using the previous algorithm.Finally,DLS is used to fit hyperbolas and find their vertices,thus achieving accurate root localization.Experimental results show that its noise reduction effect is better than similar methods.Furthermore,the maximum error distance of root localization is 1.65 cm,and the average error distance is 1.14 cm,which indicates that it can achieve an accurate prediction of root position.(3)To address the problem of difficulty in accurately predicting root radius from GPR B-Scan data,this paper proposes a prediction algorithm for tree root radius and depth based on GPR and convolutional neural network using GPR A-Scan data as the research object.First,the A-Scan data required in the real data can be obtained through the first two algorithms.Then,the GPR A-Scan data is imported into an attention module to redistribute weights of feature information and highlight the impact of critical features on the model.Then,feature information is extracted by convolutional layers.Finally,through fully connected layers,the local features learned by previous convolutional layers are integrated into global features to predict root radius and depth.Experimental results show that the proposed algorithm can accurately predict root radius and depth.On real data,the maximum error of radius prediction is 1.56 mm,the maximum error of depth prediction is 9.90 mm,and the total average relative error is 5.83%.