Subsurface Target Identification For Full-polarimetric Ground Penetrating Radar Based On Machine Learning

Ground penetrating radar(GPR)is an exploration technology that uses electromagnetic waves to characterize underground structures.The traditional GPR generally only uses the frequency,amplitude,phase information and the length and width of the targets to identify them.However,these single polarimetric GPRs have very limited information of targets.Many information related to the morphology and structure of the target cannot be collected and recorded,which leads to a low accuracy of target recognition.In order to obtain more comprehensive information about the targets,Full-polarimetric GPR was developed and used to classify and identify the targets.Full-polarimetric GPR uses four different antenna combination methods to detect the targets,which can obtain 4 times the amount of information compared to the traditional GPR.By changing the placement of the transmitting antenna and the receiving antenna,not only can it obtains the frequency,amplitude,phase and other information that can be obtained by traditional GPR,but also the polarimetric properties of the targets can be obtained.Using these polarimetric properties,we can detect the target more accurately.However,how to classify these polarization properties more accurately is a problem that needs further research.Machine learning is a new technology developed in recent years,and it has been widely used in various disciplines.Machine learning is mainly to design some automatic learning algorithms for extracting the characteristic attributes of samples from existing data,and use these attributes to predict new unknown data.The characteristics of machine learning are that only a small amount of human operation is required to enable the machine to make automatic and fast predictions of newunknown data,with high computing efficiency and accuracy,and it can analyze and predict large amounts of data.This study mainly uses machine learning algorithms to classify and recognize the underground targets data of full-polarimetric GPR.This paper is mainly introduced from the following aspects:1.The data acquisition work of full-polarimetric GPR is introduced.Including the use of the full-polarimetric GPR data acquisition system for data acquisition in the laboratory and outdoor use of commercial ground penetrating radar for field data acquisition.2.The basic theory of classical H-Alpha decomposition and its classification template are introduced,and then tested with full-polarimetric GPR data of different targets measured in the laboratory.The limitations of the application of the classic H-Alpha classification template in the full-polarimetric GPR and its causes are analyzed,and it is proposed that machine learning methods can be used to modify the classic H-Alpha classification template.3.The method of combining support vector machine(SVM)and H-Alpha decomposition is developed and applied to the classification of underground target of full-polarimetric GPR,which improves the accuracy of classification.4.The target classification method of particle center supported plane(PCSP)is established and applied to the classification and recognition of underground targets of full-polarimetric GPR.In terms of classification accuracy,computation time and computation memory,the particle center supported plane(PCSP)method is compared with the support vector machine(SVM)method and the classic H-Alpha classification template to compare the classification effect.5.The particle center supported plane(PCSP)method is applied to underground pipe detection and mine detection,and the feasibility of the method is verified by experiments.