Computational methods for processing ground penetrating radar data

摘要

The aim of this work was to investigate signal processing and analysistechniques for Ground Penetrating Radar (GPR) and its use in civilengineering and construction industry. GPR is the general term applied totechniques which employ radio waves, typically in the Mega Hertz and GigaHertz range, to map structures and features buried in the ground or in manmadestructures. GPR measurements can suffer from large amount of noise.This is primarily caused by interference from other radio-wave-emittingdevices (e.g., cell phones, radios, etc.) that are present in the surroundingarea of the GPR system during data collection. In addition to noise, presenceof clutter – reflections from other non-target objects buried underground inthe vicinity of the target can make GPR measurement difficult to understandand interpret, even for the skilled human, GPR analysts.This thesis is concerned with the improvements and processes that canbe applied to GPR data in order to enhance target detection andcharacterisation process particularly with multivariate signal processingtechniques. Those primarily include Principal Component Analysis (PCA)and Independent Component Analysis (ICA). Both techniques have beeninvestigated, implemented and compared regarding their abilities to separatethe target originating signals from the noise and clutter type signals present in the data. Combination of PCA and ICA (SVDPICA) and two-dimensionalPCA (2DPCA) are the specific approaches adopted and further developed inthis work. Ability of those methods to reduce the amount of clutter andunwanted signals present in GPR data have been investigated and reportedin this thesis, suggesting that their use in automated analysis of GPR imagesis a possibility.Further analysis carried out in this work concentrated on analysing theperformance of developed multivariate signal processing techniques and atthe same time investigating the possibility of identifying and characterisingthe features of interest in pre-processed GPR images. The driving ideabehind this part of work was to extract the resonant modes present in theindividual traces of each GPR image and to use properties of those poles tocharacterise target. Three related but different methods have beenimplemented and applied in this work – Extended Prony, Linear PredictionSingular Value Decomposition and Matrix Pencil methods. In addition tothese approaches, PCA technique has been used to reduce dimensionality ofextracted traces and to compare signals measured in various experimentalsetups. Performance analysis shows that Matrix Pencil offers the bestresults.