Inversion Research Of Ground Penetrating Radar Based On Deep Learning

In order to better meet the needs of actual exploration engineering investigations,the inversion of ground penetrating radar B-scan images into dielectric models is intuitive and clear,which can greatly improve the efficiency and accuracy of engineering exploration.Traditional inversion methods have slow calculation speed,low accuracy and reliability,and existing deep learning-based ground penetrating radar inversion methods have incomplete feature extraction and poor spatial correspondence.In view of the above problems,the inversion research of ground penetrating radar B-scan images based on convolutional neural network is carried out.The specific research contents are as follows:(1)Improve the forward modeling program based on the finite element method,simulate the designed model scene,summarize and analyze the characteristics and laws of the ground penetrating radar B-scan image.And use this program to collect the ground penetrating radar data of single anomaly body and multi-parameter combination of position,size,shape and different permittivity,and generate B-scan images to build a sample database.There are 10 models with single anomaly and multiple anomalies and multi-parameter combinations,as well as 3 background media and 4 target media.(2)Based on the design of deep learning technology,combined with the characteristics of the ground radar B-scan image,the inversion network GPR-DSConv Net is designed.The SE attention mechanism and the ASPP architecture are designed in the network to extract effective features,so as to solve the feature of depth-direction attenuation in the ground penetrating radar signal.Multi-level feature fusion of global feature encoding and local feature encoding is designed to ensure the spatial correspondence between GPR data and dielectric models.Using an encoder-decoder architecture,the dielectric model image is reconstructed from the compressed features.(3)The inversion network GPR-DSConv Net is constructed and designed based on the Pytorch deep learning framework.Based on the sample database,the training set,the test set and the verification set are divided and trained for multiple times.The training results show that the designed inversion network GPR-DSConv Net has an excellent inversion effect and can clearly invert the position,size and shape of the target abnormal body.The total inversion accuracy reaches 89.19%,MIo U is 78.78%,MPA is 85.53%.Among them,the effect of single abnormal body inversion is better than that of mixed abnormal body inversion,and the outline of abnormal body is more clear and regular.Ablation experiments are designed to verify the improvement of the inversion network,and the inversion results of different frequencies,dielectric constants and shapes are randomly selected to compare and analyze,which proves the effectiveness of the inversion network GPR-DSConv Net.The inversion results with different frequencies,different permittivity and different shapes are randomly selected and compared,and the effectiveness of the GPR-DSConv Net is proved.The inversion results of GPR-DSConv Net prove the inversion efficiency and relevance of GPR-DSConv Net in a field exploration site.