Ground Penetrating Radar Forward Modeling Based On Parallel Computing And Intelligent Identification Of Typical Karst Depression Features

Highway construction in southwest China faced a large number of karst geological problems in Karst depressions.The potential collapse of soil caves and karst caves has brought great harm to the construction and operation of expressways.It is of great significance for the safe construction and operation of Expressway to find out the location,type,filling situation of soil caves and karst caves in karst depression and deal with them pertinently.As a fast,nondestructive and efficient detection method,ground penetrating radar(GPR)has become the main means of geological anomaly exploration in typical karst depressions.With the rapid increase of the amount of ground penetrating radar(GPR)detection data,it is impossible to meet the needs of rapid and accurate discrimination of GPR images by manual identification.Based on the background of roadbed exploration in karst depression area of Yibi Expressway,in order to realize the fast and intelligent recognition of typical geological radar images in Karst depression,this paper divides the main geological models of karst depression by means of field investigation and theoretical analysis,develops a fast geological radar forward simulation software based on parallel calculation,and simulates different geologies based on geological model of karst depression.The main achievements are as followed(1)According to the factors of karst depression geology origin and drainage mode,the geological model of karst depression is divided in detail,and the relationship between relative dielectric constant of clay layer and depth under complex geological conditions is established.(2)In order to improve the accuracy of forward simulation of ground penetrating radar(GPR),the high-order symplectic difference form of two-dimensional finite difference time domain(FDTD)is deduced;the high-order symplectic difference form of absorption boundary of convolution perfect matching layer is deduced;the numerical stability and dispersion conditions in forward simulation of GPR are deduced;and the function form of Rayleigh wavelet transmitter is deduced.(3)The finite difference time domain(FDTD)algorithm is accelerated by using GPU parallel programming.FastGPRvl.0,a two-dimensional high-order symplectic differential forward simulation software for ground penetrating radar(GPRv1.0),and a forward simulation drawing software for ground penetrating radar(GPRv1.0)are developed.Under the same calculation accuracy,the calculation speed is increased by nearly 50 times.(4)FasGPRv1.0 is used to simulate various factors such as relative dielectric constant modes of different overburden,different span-height ratios and different geological bodies.A total of 1680 geological radar forward simulation images are generated for the development and training of intelligent recognition algorithm of geological radar images.(5)A variety of image data enhancement methods are used to enhance the data set of typical geological hazard radar forward simulation in Karst depression.The SSD target detection algorithm is implemented based on Tensorflow depth convolution neural network framework;the algorithm is trained on the forward simulation image set of geological radar,and the average accuracy of the algorithm is close to 78%on the test set;the relevant verification work is done on the field acquisition of geological radar images to verify the reliability of the depth learning algorithm trained by forward simulation data set.