A Deep Learning Method For Ground-penetrating Radar B-scan Data Detection At Yaohe Plateau Site

Traditional site detection involves a large number of researchers using Luoyang shovels and other tools to conduct a thorough search at regular intervals in predetermined locations.Nevertheless,this method has the disadvantages of being time-consuming,expensive,and limited in its ability to cover a large area.To promote the archaeological process of Yaoheyuan,the Yaoheyuan Project of Inner Mongolia University in Ningxia uses Ground Penetrating Radar(GPR)-based scientific and technological archaeological methods to outline the distribution range of Yaoheyuan sites and the burial depth of cultural relics in a systematic manner.Based on the Yaoheyuan Project of Inner Mongolia University,this paper aims to propose a deep learning-based hyperbolic detection method of ground penetrating radar B-scan data,addressing the issues that traditional ground penetrating radar data processing relies excessively on professionals’ knowledge and the inefficient process of professionals’ analysis of radar data.The main focus of this paper is as follows:1.This study addresses the problem of data collection in the site area and the need for a substantial number of data sets in deep learning training networks.The researchers developed a GPR forward simulation model with FDTD theory and GPRSIM software to overcome these challenges.The model was utilized to investigate the imaging features of buried targets and generate a significant number of forward simulation data sets.An enhanced R-CNN feature extraction model has been developed based on the characteristics of small objects found in subterranean ruins.Additionally,a target detection network has been established using simulated data sets.2.In this paper,four distinct experiments involving the embedding of concrete steel pipes are designed to address the problem of selecting antennas suitable for the frequency and detection depth in site data acquisition.The detection process involves the utilization of antennas operating at four distinct frequencies,which enables the identification of four distinct clusters of steel pipes.It is indicated from the findings that within the soil environment of the designated location,the 200 MHz,400 MHz,and 900 MHz antennas exhibit comparable detection depths for small subterranean entities,with each measuring approximately 1 meter.Based on comparative analysis,the optimal image resolution is achieved by acquiring a 400 MHz antenna.3.This study uses various techniques,including filtering and gain adjustments to process the images obtained through ground-penetrating radar using GPRSLICE.It aims to address the challenges posed by the uneven distribution of soil media and the interference of underground clutter noise in the target environment of the site area,which can affect the quality of the collected images.The model for detecting targets is developed using processed data obtained from measurements.This paper presents the design of three backbone networks,namely S-Res Net50,C-Res Net50,and E-Res Net50,with the incorporation of an attention mechanism.Additionally,the paper introduces the concept of transfer learning.The simulation data set’s established training weight serves as the basis for the model’s pre-training weight.The test set exhibits accuracies of 96.53%,97.19%,and 98.01%,respectively.