Dans ce projet, nous étudions la reconstruction de milieux terrestres souterrains.L’imagerie sismique est traitée avec un problème de minimisation itérative àgrande échelle, et nous utilisons la méthode de l’inversion des formes d’ondes(Full Waveform Inversion, FWI method). La reconstruction est basée sur desmesures d’ondes sismiques, car ces ondes sont caractérisées par le milieu danslequel elles se propagent. Tout d’abord, nous présentons les méthodesnumériques qui sont nécessaires pour prendre en compte l’hétérogénéité etl’anisotropie de la Terre. Ici, nous travaillons avec les solutions harmoniques deséquations des ondes, donc dans le domaine fréquentiel. Nous détaillons leséquations et l’approche numérique mises en place pour résoudre le problèmed’onde.Le problème inverse est établi afin de reconstruire les propriétés du milieu. Ils’agit d’un problème non-linéaire et mal posé, pour lequel nous disposons de peude données. Cependant, nous pouvons montrer une stabilité de type Lipschitzpour le problème inverse associé avec l’équation de Helmholtz, en considérantdes modèles représentés par des constantes par morceaux. Nous explicitons laborne inférieure et supérieure pour la constante de stabilité, qui nous permetd’obtenir une caractérisation de la stabilité en fonction de la fréquence et del’échelle. Nous revoyons ensuite le problème de minimisation associé à lareconstruction en sismique. La méthode de Newton apparaît comme naturelle,mais peut être difficilement accessible, dû au coup de calcul de la Hessienne.Nous présentons une comparaison des méthodes pour proposer un compromisentre temps de calcul et précision. Nous étudions la convergence de l’algorithme,en fonction de la géométrie du sous-sol, la fréquence et la paramétrisation. Celanous permet en particulier de quantifier la progression en fréquence, en estimantla taille du rayon de convergence de l’espace des solutions admissibles.A partir de l’étude de la stabilité et de la convergence, l’algorithme deminimisation itérative est conduit en faisant progresser la fréquence et l’échellesimultanément. Nous présentons des exemples en deux et trois dimensions, etillustrons l’incorporation d’atténuation et la considération de milieux anisotropes.Finalement, nous étudions le cas de reconstruction avec accès aux données deCauchy, motivé par les dual sensors développés en sismique. Cela nous permetde définir une nouvelle fonction coût, qui permet de prometteuses perspectivesavec un besoin minimal quant aux informations sur l’acquisition. / In this project, we investigate the recovery of subsurface Earth parameters. Weconsider the seismic imaging as a large scale iterative minimization problem, anddeploy the Full Waveform Inversion (FWI) method, for which several aspects mustbe treated. The reconstruction is based on the wave equations because thecharacteristics of the measurements indicate the nature of the medium in whichthe waves propagate. First, the natural heterogeneity and anisotropy of the Earthrequire numerical methods that are adapted and efficient to solve the wavepropagation problem. In this study, we have decided to work with the harmonicformulation, i.e., in the frequency domain. Therefore, we detail the mathematicalequations involved and the numerical discretization used to solve the waveequations in large scale situations.The inverse problem is then established in order to frame the seismic imaging. Itis a nonlinear and ill-posed inverse problem by nature, due to the limitedavailable data, and the complexity of the subsurface characterization. However,we obtain a conditional Lipschitz-type stability in the case of piecewise constantmodel representation. We derive the lower and upper bound for the underlyingstability constant, which allows us to quantify the stability with frequency andscale. It is of great use for the underlying optimization algorithm involved to solvethe seismic problem. We review the foundations of iterative optimizationtechniques and provide the different methods that we have used in this project.The Newton method, due to the numerical cost of inverting the Hessian, may notalways be accessible. We propose some comparisons to identify the benefits ofusing the Hessian, in order to study what would be an appropriate procedureregarding the accuracy and time. We study the convergence of the iterativeminimization method, depending on different aspects such as the geometry ofthe subsurface, the frequency, and the parametrization. In particular, we quantifythe frequency progression, from the point of view of optimization, by showinghow the size of the basin of attraction evolves with frequency. Following the convergence and stability analysis of the problem, the iterativeminimization algorithm is conducted via a multi-level scheme where frequencyand scale progress simultaneously. We perform a collection of experiments,including acoustic and elastic media, in two and three dimensions. Theperspectives of attenuation and anisotropic reconstructions are also introduced.Finally, we study the case of Cauchy data, motivated by the dual sensors devicesthat are developed in the geophysical industry. We derive a novel cost function,which arises from the stability analysis of the problem. It allows elegantperspectives where no prior information on the acquisition set is required.

Les premiers mètres à centaines de mètres de la proche surface terrestre sont le siège de processus naturels dont la compréhension requiert une caractérisation fine de la subsurface, via une estimation quantifiée de ses paramètres. Le géoradar est un outil de prospection indirecte à même d'ausculter les milieux naturels et d'en estimer les propriétés électriques (permittivité et conductivité). Basé sur la propagation d'ondes électromagnétiques à des fréquences allant du MHz à quelques GHz, le géoradar est utilisé à des échelles et pour des applications variées concernant la géologie, l'hydrologie ou le génie civil. Dans ce travail de thèse, je propose une méthode d'imagerie quantitative des propriétés électriques sur des sections 2D de la subsurface, à partir de données radar acquises à la surface du sol. La technique mise en oeuvre est l'inversion des formes d'ondes, qui utilise l'intégralité du champ d'ondes enregistré.Dans une première partie, je présente les principes physiques et l'outil de modélisation numérique utilisés pour simuler la propagation des ondes électromagnétiques dans les milieux hétérogènes à deux dimensions. Pour cela, un algorithme de différences finies en domaine fréquentiel développé dans le cadre des ondes visco-acoustiques est adapté au problème électromagnétique 2D grâce à une analogie mathématique.Dans une deuxième partie, le problème d'imagerie est formulé sous la forme d'une optimisation multi-paramètre puis résolu avec l'algorithme de quasi-Newton L-BFGS. Cet algorithme permet d'estimer l'effet de la matrice Hessienne, dont le rôle est crucial pour la reconstruction de paramètres de différents types comme la permittivité et la conductivité. Des tests numériques montrent toutefois que l'algorithme reste sensible aux échelles utilisées pour définir ces paramètres. Dans un exemple synthétique représentatif de la proche surface, il est cependant possible d'obtenir des cartes 2D de permittivité et de conductivité à partir de données de surface, en faisant intervenir des facteurs d'échelle et de régularisation visant à contraindre les paramètres auxquelles l'inversion est la moins sensible. Ces facteurs peuvent être déterminés en analysant la qualité de l'ajustement aux données, sans hypothèse a priori autre que la contrainte de lissage introduite par la régularisation.Dans une dernière partie, la méthode d'imagerie est confrontée à deux jeux de données réelles. Dans un premier temps, l'examen de données expérimentales permet de tester la précision des simulations numériques vis-à-vis de mesures effectuées en environnement contrôlé. La connaissance des cibles à imager permet en outre de valider la méthodologie proposée pour l'imagerie multiparamètre dans des conditions très favorables puisqu'il est possible de calibrer le signal source et de considérer l'espace libre environnant les cibles comme modèle initial pour l'inversion.Dans un deuxième temps, j'envisage le traitement d'un jeu de données radar multi-offsets acquises au sein d'un massif calcaire. L'interprétation de ces données est rendue beaucoup plus difficile par la complexité du milieu géologique environnant, ainsi que par la méconnaissance des caractéristiques précises des antennes utilisées. L'application de la méthode d'inversion des formes d'ondes à ces données requiert donc une étape préliminaire impliquant une analyse de vitesse plus classique, basée sur les arrivées directes et réfléchies, et des simulations numériques dans des modèles hypothétiques à même d'expliquer une partie des données. L'estimation du signal source est effectuée à partir d'arrivées sélectionnées, simultanément avec des valeurs moyennes de conductivité et de hauteur d'antennes de façon à reproduire au mieux les amplitudes observées. Un premier essai d'inversion montre que l'algorithme est capable d'expliquer les données dans la gamme de fréquences considérée et de reconstruire une ébauche des principaux réflecteurs. / The quantitative characterization of the shallow subsurface of the Earth is a critical issue for many environmental and societal challenges. Ground penetrating radar (GPR) is a geophysical method based on the propagation of electromagnetic waves for the prospection of the near subsurface. With central frequencies between 10~MHz and a few GHz, GPR covers a wide range of applications in geology, hydrology and civil engineering. GPR data are sensitive to variations in the electrical properties of the medium which can be related, for instance, to its water content and bring valuable information on hydrological processes. In this work, I develop a quantitative imaging method for the reconstruction of 2D distributions of permittivity and conductivity from GPR data acquired from the ground surface. The method makes use of the full waveform inversion technique (FWI), originating from seismic exploration, which exploits the entire recorded radargrams and has been proved successful in crosshole GPR applications.In a first time, I present the numerical forward modelling used to simulate the propagation of electromagnetic waves in 2D heterogeneous media and generate the synthetic GPR data that are compared to the recorded radargrams in the inversion process. A frequency-domain finite-difference algorithm originally developed in the visco-acoustic approximation is adapted to the electromagnetic problem in 2D via an acoustic-electromagnetic mathematical analogy.In a second time, the inversion scheme is formulated as a fully multiparameter optimization problem which is solved with the quasi-Newton L-BFGS algorithm. In this formulation, the effect of an approximate inverse Hessian is expected to mitigate the trade-off between the impact of permittivity and conductivity on the data. However, numerical tests on a synthetic benchmark of the literature display a large sensitivity of the method with respect to parameter scaling, showing the limits of the L-BFGS approximation. On a realistic subsurface benchmark with surface-to-surface configuration, it has been shown possible to ally parameter scaling and regularization to reconstruct 2D images of permittivity and conductivity without a priori assumptions.Finally, the imaging method is confronted to two real data sets. The consideration of laboratory-controlled data validates the proposed workflow for multiparameter imaging, as well as the accuracy of the numerical forward solutions. The application to on-ground GPR data acquired in a limestone massif is more challenging and necessitates a thorough investigation involving classical processing techniques and forward simulations. Starting permittivity models are derived from the velocity analysis of the direct arrivals and of the reflected events. The estimation of the source signature is performed together with an evaluation of an average conductivity value and of the unknown antenna height. In spite of this procedure, synthetic data do not reproduce the observed amplitudes, suggesting an effect of the radiation pattern of the shielded antennae. In preliminary tests, the inversion succeeds in fitting the data in the considered frequency range and can reconstruct reflectors from a smooth starting model.

The primary focus of this thesis is to examine the detailed seismic structure of the northern Cascadia margin, including the Cascadia basin, the deformation front and the continental shelf. The results of this study are contributing towards understanding sediment deformation and tectonics on this margin. They also have important implications for exploration of hydrocarbons (oil and gas) and natural hazards (submarine landslides, earthquakes, tsunamis, and climate change). The first part of this thesis focuses on the role of gas hydrate in slope failure observed from multibeam bathymetry data on a frontal ridge near the deformation front off Vancouver Island margin using active-source ocean bottom seismometer (OBS) data collected in 2010. Volume estimates (∼ 0.33 km^3) of the slides observed on this margin indicate that these are capable of generating large (∼ 1 − 2 m) tsunamis. Velocity models from travel time inversion of wide angle reflections and refractions recorded on OBSs and vertical incidence single channel seismic (SCS) data were used to estimate gas hydrate concentrations using effective medium modeling. Results indicate a shallow high velocity hydrate layer with a velocity of 2.0 − 2.1 km/s that corresponds to a hydrate concentration of 40% at a depth of 100 m, and a bottom simulating reflector (BSR) at a depth of 265 − 275 m beneath the seafloor (mbsf). These are comparable to drilling results on an adjacent frontal ridge. Margin perpendicular normal faults that extend down to BSR depth were also observed on SCS and bathymetric data, two of which coincide with the sidewalls of the slump indicating that the lateral extent of the slump is controlled by these faults. Analysis of bathymetric data indicates, for the first time, that the glide plane occurs at the same depth as the shallow high velocity layer (100±10 mbsf). In contrast, the glide plane coincides with the depth of the BSR on an adjacent frontal ridge. In either case, our results suggest that the contrast in sediments strengthened by hydrates and overlying or underlying sediments where there is no hydrate is what causing the slope failure on this margin. The second part of this dissertation focuses on obtaining the detailed structure of the Cascadia basin and frontal ridge region using mirror imaging of few widely spaced OBS data. Using only a small airgun source (120 cu. in.), our results indicate structures that were previously not observed on the northern Cascadia margin. Specifically, OBS migration results show dual-vergence structure, which could be related to horizontal compression associated with subduction and low basal shear stress resulting from over-pressure. Understanding the physical and mechanical properties of the basal layer has important implications for understanding earthquakes on this margin. The OBS migrated image also clearly shows the continuity of reflectors which enabled the identification of thrust faults, and also shows the top of the igneous oceanic crust at 5−6 km beneath the seafloor, which were not possible to identify in single-channel and low-fold multi-channel seismic (MCS) data. The last part of this thesis focuses on obtaining detailed seismic structure of the Vancouver Island continental shelf from MCS data using frequency domain viscoacoustic full waveform inversion, which is first of its kind on this margin. Anelastic velocity and attenuation models, derived in this study to subseafloor depths of ∼ 2 km, are useful in understanding the deformation within the Tofino basin sediments, the nature of basement structures and their relationship with underlying accreted terranes such as the Crescent and the Pacific Rim terranes. Specifically, our results indicate a low-velocity zone (LVZ) with a contrast of 200 m/s within the Tofino basin sediment section at a depth 600 − 1000 mbsf over a lateral distance of 10 km. This LVZ is associated with high attenuation values (0.015 − 0.02) and could be a result of over pressured sediments or lithology changes associated with a high porosity layer in this potential hydrocarbon environment. Shallow high velocities of 4 − 5 km/s are observed in the mid-shelf region at depths > 1.5 km, which is interpreted as the shallowest occurrence of the Eocene volcanic Crescent terrane. The sediment velocities sharply increase about 10 km west of Vancouver Island, which probably corresponds to the underlying transition to the Mesozoic marine sedimentary Pacific Rim terrane. High attenuation values of 0.03 − 0.06 are observed at depths > 1 km, which probably corresponds to increased clay content and the presence of mineralized fluids. / Graduate / 0373 / 0372 / 0605 / subbarao@uvic.ca

Humanitarian clearance of minefields is most often carried out by hand, conventionally using a a metal detector and a probe. Detection is a very slow process, as every piece of detected metal must treated as if it were a landmine and carefully probed and excavated, while many of them are not. The process can be safely sped up by use of Ground-Penetrating Radar (GPR) to image the subsurface, to verify metal detection results and safely ignore any objects which could not possibly be a landmine. In this thesis, we explore the possibility of using Full Wave Inversion (FWI) to improve GPR imaging for landmine detection. Posing the imaging task as FWI means solving the large-scale, non-linear and ill-posed optimisation problem of determining the physical parameters of the subsurface (such as electrical permittivity) which would best reproduce the data. This thesis begins by giving an overview of all the mathematical and implementational aspects of FWI, so as to provide an informative text for both mathematicians (perhaps already familiar with other inverse problems) wanting to contribute to the mine detection problem, as well as a wider engineering audience (perhaps already working on GPR or mine detection) interested in the mathematical study of inverse problems and FWI.We present the first numerical 3D FWI results for GPR, and consider only surface measurements from small-scale arrays as these are suitable for our application. The FWI problem requires an accurate forward model to simulate GPR data, for which we use a hybrid finite-element boundary-integral solver utilising first order curl-conforming N\'d\'{e}lec (edge) elements. We present a novel `line search' type algorithm which prioritises inversion of some target parameters in a region of interest (ROI), with the update outside of the area defined implicitly as a function of the target parameters. This is particularly applicable to the mine detection problem, in which we wish to know more about some detected metallic objects, but are not interested in the surrounding medium. We may need to resolve the surrounding area though, in order to account for the target being obscured and multiple scattering in a highly cluttered subsurface. We focus particularly on spatial sensitivity of the inverse problem, using both a singular value decomposition to analyse the Jacobian matrix, as well as an asymptotic expansion involving polarization tensors describing the perturbation of electric field due to small objects. The latter allows us to extend the current theory of sensitivity in for acoustic FWI, based on the Born approximation, to better understand how polarization plays a role in the 3D electromagnetic inverse problem. Based on this asymptotic approximation, we derive a novel approximation to the diagonals of the Hessian matrix which can be used to pre-condition the GPR FWI problem.