Вы находитесь на странице: 1из 10

GEOPHYSICS. VOL. 77, NO. 1 (JANUARY-FEBRUARY 2012); P. H9H18, 8 FIGS. 10.1190/GEO2011-0147.

Focused inversion of vertical radar profile (VRP) traveltime data

Giulio Vignoli1, Rita Deiana2, and Giorgio Cassiani2

ABSTRACT
The reconstruction of the GPR velocity vertical profile from vertical radar profile (VRP) traveltime data is a problem with a finite number of measurements and imprecise data, analogous to similar seismic techniques, such as the shallow down-hole test used for S-wave velocity profiling or the vertical seismic profiling (VSP) commonly used in deeper exploration. The uncertainty in data accuracy and the error amplification inherent in deriving velocity estimates from gradients of arrival times make this an example of an ill-posed inverse problem. In the framework of Tikhonov regularization theory, ill-posedness can be tackled by introducing a regularizing functional (stabilizer). The role of this functional is to stabilize the numerical solution by incorporating the appropriate a priori assumptions about the geometrical and/or physical properties of the solution. One of these assumptions could be the existence of sharp boundaries

separating rocks with different physical properties. We apply a method based on the minimum support stabilizer to the VRP traveltime inverse problem. This stabilizer makes it possible to produce more accurate profiles of geological targets with compact structure. We compare more traditional inversion results with our proposed compact reconstructions. Using synthetic examples, we demonstrate that the minimum support stabilizer allows an improved recovery of the profile shape and velocity values of blocky targets. We also study the stabilizer behavior with respect to different noise levels and different choices of the reference model. The proposed approach is then applied to real cases where VPRs have been used to derive moisture content profiles as a function of depth. In these real cases, the derived sharper profiles are consistent with other evidence, such as GPR zero-offset profiles, GPR reflections and known locations of the water table.

INTRODUCTION
The use of near-surface geophysics as a contribution to the solution of hydrological and environmental problems has become of general acceptance. The corresponding applications are generally referred to under the general name of hydrogeophysics (Rubin and Hubbard, 2005; Vereecken et al., 2006). Several techniques have been extensively used for the hydrogeophysical characterization of the soil and subsoil. Among them, the most popular are electrical resistivity tomography (ERT e.g., Binley and Kemna, 2005) and ground-penetrating radar (GPR e.g., Davis and Annan, 1989; Annan, 2005). GPR is particularly suitable for vadose zone characterization and monitoring (Cassiani et al., 2006) as electromagnetic wave propagation velocity is directly related to the soil volumetric moisture content (e.g., Greaves et al., 1996; Strobbia and Cassiani, 2007; Wijewardana and Galagedara, 2010) and therefore

porosity below the water table (e.g., Turesson, 2006), provided that suitable constitutive relationships are established (e.g., Topp et al., 1980; Roth et al., 1990; Brovelli and Cassiani, 2010). ERT and GPR have great advantages when used in borehole configuration because of the dramatic increase in resolution that can be achieved at depth. In addition, borehole GPR is superior to surface-to-surface GPR because (1) penetration is enhanced by avoiding the shallow, usually electrically conductive, soil layers, and (2) measurement of velocity is simplified by using direct waves between transmitter and receiver. As a consequence, a number of hydrogeophysical studies have adopted crosshole GPR (e.g., Hubbard et al., 1997; Alumbaugh et al., 2002; Binley et al., 2002; Binley and Beven, 2003; Rucker and Ferr, 2004; Tronicke et al., 2004; Cassiani and Binley, 2005; Deiana et al., 2007, 2008; Looms et al., 2008a, b). However, these applications require two boreholes at a distance not larger than a few meters, a situation rarely available at sites of

Manuscript received by the Editor 18 April 2011; revised manuscript received 9 September 2011; published online 3 February 2012. 1 King Fahd University of Petroleum & Minerals, Earth Sciences Department, Dhahran, Saudi Arabia. E-mail: giulio.vignoli@gmail.com. 2 Universit di Padova, Dipartimento di Geoscienze, Padova, Italy. E-mail: rita.deiana@unipd.it; giorgio.cassiani@unipd.it. 2012 Society of Exploration Geophysicists. All rights reserved. H9

Downloaded 05 Feb 2012 to 147.162.110.99. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

H10

Vignoli et al. reference model. Furthermore, we discuss the results concerning real VRP data collected for hydrogeophysical purposes at a few selected test sites and we compare these results with independent evidence from GPR zero-offset profiles, VRP reflections and water table information. The comparison between focused and nonfocused stabilizers in hydrogeophysical applications is particularly significant because, quite often, abrupt variations of the moisture content (and consequently of the velocity) may occur. In those cases, a blind application of a traditional MN-like stabilizer would generate probably oversmoothed, misleading profiles.

practical interest. As a consequence, being a single-hole approach, vertical radar profile (VRP) is often advantageous (e.g., Cassiani et al., 2004; Tronicke and Knoll, 2005; Clement and Knoll, 2006; Bniger et al., 2006; Cassiani et al., 2008; Addison et al., 2009). Even if several differences exist (Tronicke and Knoll, 2005), VRP is for a large extent analogous to vertical seismic profiling (VSP), a technique widely used in seismic exploration, both in terms of velocity estimates (Moret et al., 2004) and reflection analysis (Hardage, 2000). The velocity profile and the location of impedance contrasts are the information that can be extracted from VRP data. Given the strong link between velocity and impedance, one would expect that these two pieces of information are in fair agreement, and reflection locations are consistent with strong contrasts in propagation velocity. The observed traveltime at a specific depth is the definite integral of the vertical slowness (i.e., the inverse of the velocity) from the surface to that depth; in principle, by using the fundamental theorem of calculus, the slowness can be calculated by the derivative of the traveltime. Unfortunately, the identification of the GPR velocity profile from VRP traveltime data is a problem with a finite number of measurements, and imprecise data. The uncertainty in data accuracy and the error amplification inherent in deriving velocity estimates from gradients of arrival times make this an ill-posed inverse problem. A general strategy to overcome ill-posedness is to include the available a priori information in the process of solution search (e.g., Jackson, 1979; Lelievre and al., 2009). Thus, the selected solution is the one compatible with: (1) the data, and (2) our assumptions about the investigated system. For example, if we know (e.g., from other geological and/or geophysical investigations) that the soil is layered and we expect a sharp velocity transition between layers, we should use this information to constrain the solution. In the framework of Tikhonov regularization theory (Tikhonov and Arsenin, 1977), a priori information is implemented into inversion through the inclusion of an appropriate choice of the stabilizing functional (stabilizer). The most widely used stabilizers are smoothing functionals that constrain the first or second spatial derivative of the sought parameter field. This approach is the most commonly used for VRP and VSP velocity analysis (Lizzaralde and Shift, 1999; Clement and Knoll 2006) thus leading to smooth velocity profiles that are inconsistent with the presence of strong reflectors (e.g., Cassiani et al., 2004) or of a welldefined interface of dielectric properties such as the water table. In these situations, it would be useful to search for a stable solution within the class of inverse blocky models with sharp petrophysical boundaries. There exist several alternatives to stabilizers penalizing sharp changes in properties (e.g., Ramos et al., 1999; Bertete-Aguirre et al., 2002; Farquharson, 2008). Among them, recently, the minimum support (MS) stabilizer (Last and Kubik, 1983; Portniaguine and Zhdanov, 1999) has been applied to traveltime tomography (Zhdanov et al., 2006; Ajo-Franklin et al., 2007). This stabilizer provides a focusing regularized inversion as it minimizes the area where the solution is different from a reference model, and not their mutual (l2 ) distance as done in the much more common minimum norm (MN) methods. In this paper, we present a study of focusing inversion of VRP data. To assess the potential of a minimum support inversion algorithm on VRP data, we compare the results generated by MN and MS stabilizers for a synthetic data set and we explore their behaviors with respect to different noise levels and different choices of the

THE LINEAR TRAVELTIME INVERSE PROBLEM


Within the geometric optics approximation, the traveltime r 0; rj of the radar ray from the source r 0 to the receiver rj, can be described as a function of the slowness sr, i.e., the inverse of the local velocity vr

r 0 ; rj

Z
Lr 0 ;rj

srdl;

(1)

where Lr0 ; r is the source-receiver path. In general, equation 1 is nonlinear because the raypath Lr0 ; r depends on the slowness spatial distribution sr the vector r spans the 3D space. Let us assume that the actual slowness distribution differs just for a small perturbation from a background slowness sb r: sr sb r sr. Fermats principle states that the traveltime is stationary with respect to a variation of the raypath. As a first approximation

Z sb s
Lsb s

Z sb sdl
Lsb

sb sdl; (2)

where sb s is the stationary traveltime along the actual raypath, while Lsb is the raypath calculated for the background slowness. The tomographic problem can thus be reduced to a linear problem

d Am;

(3)

where m is the model vector (consisting of the constant slowness in the M discretizing cells) and d is the data vector (consisting of the measured traveltimes i , i 1; : : : ; D), while the element Ail of the matrix A is the length of the i-th ray crossing the l-th cell. Our final goal is to invert (in a generalized sense) A to get m from d. It is worth noting that linear approximation is sometimes a rough simplification, even for small offset VRP acquisitions in horizontal layered media. Even if the ray trajectory is nearly vertical, deflections due to velocity changes can be significant (Moret et al, 2004). In general, three difficulties may arise when tackling inverse problems (Hadamard, 1902); the solution of the inverse problem, as described by equation 3: (1) may not exist; (2) may be nonunique; (3) may be unstable. Within Tikhonov regularization theory (Tikhonov and Arsenin, 1977), these difficulties are resolved by introducing the notion of Tikhonov parametric functional as the weighted sum of misfit and stabilizer functional. The mere minimization of the misfit functional solves the problem of a possible nonexistence of a solution m of equation 3. Even when a solution sensu stricto does not exist, we can be satisfied by

Downloaded 05 Feb 2012 to 147.162.110.99. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

Focused inversion of VRP traveltimes the model m0 that best fits the data, i.e., the model m0 that minimizes the distance between the observed data and the calculated response. Usually the distance in the data space is defined by the l2 norm; thus, the misfit functional is

H11

sMN

M X k1

mk maprk 2

(6)

The minimum support stabilizer sMS

m kAm dk22 : l

(4)

sMS

M X

mk maprk 2 ; mk maprk 2 e2 k1

(7)

In general, the minimizers of equation 4 could be more than one. The nonuniqueness issue comes up from these ambiguities. The stabilizer sm formalizes our a priori idea about the system, i.e., the value of sm is a measure of how much the solution m is far from our a priori idea. The smaller the sm, the greater the agreement between m and our hypothesis. While the misfit m measures the quality of our solution (in terms of adherence to observed data) in the data space, sm measures the quality (in terms of adherence to our prior information) in the model space. Thus, the minimization of the weighted sum P of these two terms

where e is named the focusing parameter. In equation 7, e is needed to avoid singularities that occur when mk maprk ; for this reason, e should satisfy the following inequality:
e2 mk maprk 2

eps, where eps is the machine floating-point

Pm; m sm

(5)

relative accuracy; i.e., eps is, by definition, the distance from 1.0 to the nearest floating-point number less than 1.0. In this paper, we set e2 eps max m mapr 2 , where maxm mapr is the largest element in the m mapr array. sMS is proportional to the number of nonzero elements of (m mapr ); in fact, a possible continuous approximation of the signum function is

implies that an appropriate approximated solution is identified within a model class with specific properties specified by sm. P is the very well known Tikhonov parametric functional. The stabilizer resolves the nonuniqueness issue by selecting, among all models compatible with the data, the one in agreement with the a priori information. Data fitting and model fitting are in trade-off, controlled by the weight. Of course, not all functionals are suitable to be stabilizers: they also should guarantee the solution stability with respect to the data variation. Briefly, the functional sm can be a stabilizer if the set of m defined by sm < q, q > 0, is a compact set (Tikhonov and Arsenin, 1977; Zhdanov, 2002). From this point of view, the regularization parameter can be seen also as an arbiter between fidelity and stability of the approximate solution. There are several criteria for the selection of the regularization parameter; in this work, we choose its value in accordance with the level of noise in the data. A possible approach to determine the level of VRP data error is proposed for instance by Cassiani et al. (2004). In summary, the formulation of a traveltime inverse problem using Tikhonov regularization reduces the original problem of equation 3 to the minimization of the Tikhonov parametric functional in equation 5. The stabilizers choice is connected to the available a priori information; in our case, we want to investigate the possibility to get compact reconstructions from VRP data.

mk signe mk p : jmk j2 e2

(8)

Consequently, sMS is approximately the sum of the squared signs of the difference between the solution parameters and their reference values (Nagarajan et al., 2006)

sMS

M X k1

signe mk maprk 2 :

(9)

In summary, sMS measures the number of cells where m is different from mapr ; thus, it measures the support of (m mapr ). The solution selected by sMS is the one characterized by the minimum support, which is the minimum area where the argument does not vanish.

PARAMETRIC FUNCTIONAL MINIMIZATION SCHEME


The regularized solution of the linearized traveltime inversion is the minimizer of Pm; (equation 5). For a large class of stabilizers, the regularizing term can be expressed as the squared norm of the weighted difference m mapr

sm kWm mapr k2

(10)

MINIMUM NORM AND MINIMUM SUPPORT STABILIZERS


In the previous section, the existence of multiple models equally consistent with the data has been discussed. Clearly, the misfit functional cannot discriminate between them. And the introduction of a priori knowledge in the inversion process is necessary. We can constrain the solution by choosing an appropriate stabilizer. There are many possible choices for the stabilizer (Vogel, 2002). In this paper, we analyze two of them:

For the MN case, the weighting matrix W is the identity matrix; for other smoothing stabilizers, e.g., Occams inversion (Constable et al., 1987), W can implement higher order differential operators. For the MS stabilizer, the expression for W is more complex because it depends on m; for the MS stabilizer, W is a diagonal matrix whose nonzero elements are

W k mk maprk 2 e2 12 :

(11)

The minimum norm stabilizer sMN which is the square l2 distance between the solution and the reference a priori model mapr :

Thus, the MS stabilizer is nonquadratic and we need an iterative approach such as a gradient-type minimization method (Fletcher, 1981). The signum function, implicit in the MS stabilizer, has a discontinuous first derivative, but its continuous approximation

Downloaded 05 Feb 2012 to 147.162.110.99. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

H12

Vignoli et al. functional reaches the minimum in few iterations (typically tens), while the MS needs more iterations. Actually, in the investigated cases, the necessary MS iterations are hundreds; thus, they seem to be of the order of magnitude of the number of model parameters. So, even if, for small data sets, like those usually collected for environmental applications, this higher computational cost turns into a hardly detectable bigger time cost, for larger data sets it might be important to develop an efficient (e.g., not a posteriori) way for the selection of the regularization parameter . This would help to reduce the overall number of required minimizations (Zhdanov et al., 2006). The weighting matrix W in equation 10 is used to discriminate among model parameters. In the MN case, no difference is made between one component of (m mapr ) and the others; in the MS case, W selects some specific components of (m mapr ) to contribute to the summation implicit in the stabilizer expression, and excludes the others. Another example of model weighting matrix could be based on the reliability of the model parameters: in traveltime tomography, velocity value of highly crossed cells is in general more constrained, and reliable, than velocity of cells with low ray coverage. The use of an appropriate W to weight differently the model parameters could conveniently take into account these differences (Zhdanov et al., 2006). Analogously, we can define a weighting matrix Wd acting on the data points, incorporating the available information on data uncertainty, e.g., defining a diagonal Wd matrix such that its elements are the inverse of the data standard errors i (i 1; : : : ; D), where D is the number of measurements. Hence, the misfit functional becomes

(equation 8) makes it possible to compute the gradient of the MS functional (equation 7) and to apply gradient-type methods. In this paper, we follow the approach described in detail in Zhdanov et al. (2006) for the direct conjugate gradient minimization of equation 5 with MS stabilizer. The elements of the steepest descent direction for the parametric functional P with MS are

lk

MS

D M XX n1 j1

Ank Anj mj dn

0 W k W k mk maprk W k mk maprk ; (12)

where W k is the first derivative of W k ,

Wk

mk maprk ; mk maprk 2 e2 32

(13)

and takes into account the dependence of the weighting matrix W on the model parameters. The first addendum in the summation of equation 12 corresponds to the application of the transpose of the tomographic matrix A to the residual vector. 0 Of course, in the MN case, because W k 1 and W k 0 for every k, the gradient direction is

lk

MN

D M XX n1 j1

Ank Anj mj dn mk maprk : (14)

We can include these expressions into a standard algorithm for the conjugate gradient minimization. Hence, at the (n 1)-th iteration, the updated model mn1 can be calculated by using the following scheme

m 2 m kWd Am Wd dk22 l
D X i1

rn Amn d; ln lMN or MS mn
(see equation 14 or equation 12)

(15) (16)

calc i 2 2 ; i i

(21)

kln k2 ; kln1 k2 ~0 l0 ; l ;

where calc is the calculated traveltime of the i-th ray. In this way, i the misfit functional m is made equal to a 2 functional. For normally distributed measurement errors, the 2 quantity should be equal to D (Taylor, 1997). So, we can use, as a discrepancy criterion (Tikhonov and Arsenin, 1977) for the selection of ,

(17) (18) (19) (20)

2 m D 1;

(22)

~ ln ln n ~n1 ; l kn ~T ln ln

kA~n k2 k~n k2 l l

where m is the minimizer of Pm; . Thus, essentially, the optimal is selected a posteriori among several different values i : after the minimization of the parametric functionals Pm; i for every i , the 0 , for which the corresponding minimizer m0 satisfies equation 22, is selected.

~ mn1 mn kn ln :

TESTS ON SYNTHETIC DATA


In this section, we discuss the performance of the MS approach when tested on a synthetic data set. We picked the traveltimes from synthetic full-waveform data generated using GprMax, a finite difference time-domain software (Giannopoulos, 2005). The structural model we investigate is shown in Figure 1. The model is characterized by several velocity inversions and different layer thicknesses. The receiver spacing along the borehole is 0.1 m, from 1 m from the model bottom to 1 m below the surface, where the transmitter is located. Figure 1 also shows the corresponding synthetic 131 traces; note the presence of well-identified reflections

From equations 1220, it is easy to see that the only implementation differences between the MN and the MS stabilizers are in the two matrices W and W 0 . For relatively small data sets, the computational cost of updating W and W 0 at each iteration is not significant. This makes the computational cost of each iteration of the MS and MN algorithms comparable in many cases. Nevertheless, in all the tests discussed in the following sections, the convergence of the MN algorithm is faster than the MS one. By using the described conjugate gradient minimization algorithm, the MN parametric

Downloaded 05 Feb 2012 to 147.162.110.99. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

Focused inversion of VRP traveltimes that appear as events with opposite slopes (e.g., Cassiani et al., 2004). The traveltime data i have been corrupted with 1% uniformly distributed noise: i pi i . p fpi g is a vector of uniformly distributed random numbers on the interval comprised between 0 and the selected noise percentage. With this approach we simulate the larger picking errors in the longer paths due to: (1) the decreasing central frequency and bandwidth caused by attenuation, and observed in real data (e.g., Cassiani et al., 2004); (2) the general lowering of the signal-to-noise-ratio for traces recorded at a deeper location (Tronicke, 2007). In all the inversions we are going to discuss, we selected, as reference models, homogeneous backgrounds. In this case, because we are performing a simplified, linearized, inversion, the raypaths are fixed and the forward modeling matrix A is calculated based on straight raypaths between the sources and receivers, disregarding any dependence on the velocity distribution. A possible way to improve the proposed inversion strategy could be the use of more sophisticated ray-tracing tools. In general, the thickness of the layers (and so the number of unknowns) used in the discretization plays an important role in the inversion. Clearly, we experienced that, for a number of cells much larger than the data (M D), the inversion result depends only on the choice of the stabilizer. Because we want to isolate and study the effect of the regularization induced by the stabilizer only, we largely over-parameterized the problem and we used 400 layers. Also from a practical point of view, even if the computational effort is slightly higher (but, with the relatively small data sets usually collected for environmental applications, the difference is hardly visible), the result is more stable as it does not depend on a parameter (the numbers of layers) otherwise difficult to control. As a first attempt, we assume to know the background velocity and we set all the component of the a priori slowness model mapr equal to 0.141 nsm. The MN and MS inversion results are presented in Figure 2a. In both cases, data are fitted assuming a uniform 0.7 ns noise level; this means that i 0.7 ns, i 1; ::; D. Note that this assumed noise level is not consistent with the level of noise added to the synthetic data set. Figure 2a

H13

clearly shows how the traditional MN inversion yields a flattened velocity profile. This result is not surprising as it is consistent with the request of a minimum distance between the solution and the reference model. The MN result is practically unable to recover the layer alternation in the top 7 m of the profile, notwithstanding the fairly large velocity variations. Moreover, the velocity value of the fast layer between 7 and 10 m is largely underestimated and the sharp transitions at its upper and lower bounds are smeared. On the other hand, the MS result can reconstruct more successfully the velocity of each layer. Doing so, the detection and location of all the velocity changes is clearly much more effective. This is true also for those discontinuities involving the shallower thin layers. Unlike MN, the MS results are quite consistent with the evidence of abrupt impedance (velocity) contrasts that give rise to the reflection events in Figure 1. Actually, with GPR field data it is desirable and often available (e.g., Alumbaugh et al., 2002; Cassiani et al., 2004) a more precise estimation of the noise content. We also tested the response of MN and MS inversions assuming different uncertainty estimates (Figure 2b and 2c). In Figure 2b we compare the results of MN

Figure 1. Synthetic model (left) and corresponding radargram (right). This model consists of six layers; from the bottom: 4 m at 0.0775 mns; 3 m at 0.14 mns; 3 m at 0.10 mns; 1 m at 0.14 mns; 1 m at 0.10 mns; 2 m at 0.14 mns. The surface is at 0 m; the receivers are located along the borehole, from 1 m to 14 m every 0.1 m, while the source is modeled on the surface.

Figure 2. Synthetic case. Comparison of the inversion results obtained with MN (turquoise line) and MS (red line). 1% uniformly distributed noise was added to the synthetic data generated from the true model (dash line). In MN and MS cases, mapri 0.14 mns, i 1; : : : ; M. For the results in panel (a) we assume a uniform data error i 0.7 ns, i 1; ::; D. For results in (b) two different uncertainty estimations are considered: i 1.0 ns in the bottom 4 m, and i 0.5 ns in the top 10 m. For panel (c) we assume to know the uncertainty distribution: for each measurement, a different i is used; the 1% noise results (MN in turquoise, MS in red) are also compared against the 3% noise reconstruction (MN in green, MS in pink).

Downloaded 05 Feb 2012 to 147.162.110.99. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

H14

Vignoli et al. any difference (larger than e) have a contribution equal to one, even for very large difference (mk maprk ). Hence, solutions differing for few layers from the a priori model are promoted, but the penalization for (large) differences is, in any case, close to one, almost disregarding the amount of the difference. Next, we investigate how robust the MN and MS results are with respect to wrong assumptions about the background velocities (Figure 3). We compared the former results obtained for mapri 0.141 nsm, i 1; : : : ; M (the maximum velocity of the true model) and i 0.7 ns, i 1; ::; D (shown in Figure 2a) with those calculated with three other background constant velocities: 0.05 mns, 0.077 mns (the minimum velocity in the true model) and 0.3 mns (the speed of light in the vacuum). For all these values, MN and MS results are very similar to the one in Figure 2a: the solutions show a very weak dependence on the choice of the reference model. We also examined the impact of actual noise level on the inversion results. We assume an approximate knowledge of the uncertainty level ( i 0.5 ns in the top 10 m and i 1.0 ns below). When a 2% uniformly distributed error is added to the synthetic data (Figure 4a), even if the features of the true model are reconstructed with less detail than before and the solution is more oscillating, the MS is still far better than the MN result. The MN stabilizer is barely capable of detecting the velocity decrease of the last layer (around 10 m), and the transition is extremely

and MS inversions performed with i 0.5 ns for the measurements collected within the shallower 10 m and i 1.0 ns for those corresponding to the deeper 4 m; in Figure 2c, we assume we have a different uncertainty value for each measurement. As one can expect, a better knowledge of the noise distribution corresponds to a more effective systems reconstruction. A simple differentiation, in terms of noise levels, between the upper and lower part of the profile (Figure 2b), allows MS inversion to reconstruct very well the boundaries of the fast deep layer and to recover quite effectively the true values of both fast inclusions close to the surface. The better the knowledge of the data, the less important the stabilizer is: Figure 2c shows that, in presence of a relatively small noise level and a precise knowledge of its distribution along the data set, the MN reconstruction can reach nearly the same quality as the MS result. Even if, in the considered examples, the MS reconstructions are generally better than the MN ones, they are characterized by a more severe, potentially misleading (small scale) fluctuating behavior. This might be imputable to the nature of the MS stabilizer. From the summation of equation 7, it is possible to see that terms, characterized by differences between the reconstructed and reference models smaller than e, have zero contribution, while terms with

Figure 3. Synthetic case. Inversion results using different choices for mapr. (a) The MN results for four different background constant velocities: 0.05 mns; 0.077 mns (the minimum velocity in the true model); 0.14 mns (the maximum velocity in the true model); 0.3 mns (the vacuum speed of light). (b) The corresponding MS results. In both cases, the impact of the reference model is largely negligible.

Figure 4. Synthetic case. Inversion results as a function of different noise levels added to the synthetic data set. In all cases, we assumed i 1.0 ns in the bottom 4 m, and i 0.5 ns in the 10 m above. MN (turquoise line) and MS (red line) inversion results are compared for data corrupted with: (a) 2%, (b) 2.5%, and (c) 3% noise. In panel (c), we show also the results obtained with a different, correct uncertainty value i for each measurement, consistent with the 3% added error (MS in pink, MN in green).

Downloaded 05 Feb 2012 to 147.162.110.99. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

Focused inversion of VRP traveltimes smeared-out; moreover, in the MN result all the velocity variations of the upper part of the model are flattened. On the other hand, the MS stabilizer with 2% noise is still able to retrieve the location of all the fast layers, even if the reconstructed velocity values are not as reliable as with a lower error level. Even for higher added noise levels (Figure 4b: 2.5%; Figure 4c: 3%), the MS stabilizer behaves better than the MN stabilizer. It is worth noting that, if we have the correct estimate of uncertainty for each data point, the MS reconstruction is very stable with respect to noise: The MS result corresponding to 3% noise (pink line in Figures 4c and 2c) is comparable with the result corresponding to 1% of noise (red line in Figure 2c). The same is not true for the MN reconstructions (green line in Figures 4c and 2c).

H15

TESTS ON FIELD DATA


Finally, we apply the MS approach to actual VRP field data. Because the experimental data were collected for hydrogeophysical purposes, the results have been converted into moisture content profiles, using the well-established approximation of the Topp et al. (1980) empirical model. As for the synthetic cases, also for the experimental data sets, we assumed a homogeneous reference model; hence, the forward modeling matrix is calculated considering straight rays between sources and receivers.

Korba site, Tunisia


The first data set comes from the Korba experimental site in the Cap Bon region, North-Eastern Tunisia, where monitoring activities were recently conducted as part of the joint research efforts of the EU FP7 CLIMB project (Ludwig et al., 2010). The Korba aquifer is housed in Plio-Quaternary loose deposits and constitutes a shallow, easy accessible water resource in a region where a semi-arid Mediterranean climate prevails. However, overexploitation of this coastal aquifer has been causing seawater intrusion and consequent groundwater quality deterioration (Paniconi et al., 2001; Kouzana et al., 2009; Kerrou et al., 2010). Reclaimed water from the Korba wastewater treatment plant is being used in the artificial recharge pilot test site of Korba El Mida, where several boreholes have

been installed to monitor water table elevation and groundwater quality, upstream and downstream of the three infiltration ponds. In June 2010, we collected VRP data on several of the existing boreholes. Here, we present only the VRP data from borehole four because the presence of borehole three only 3.8 m away allowed us also to collect a GPR zero-offset profile (ZOP see, e.g., Binley et al., 2002; Cassiani and Binley, 2005; Deiana et al., 2007, 2008) between boreholes three and four, and a comparison against the VRP inversion results is possible (Figure 5). Unfortunately, borehole three is only about 6 m deep, so the ZOP data are only available to that depth. VRP and ZOP data were collected using a Systems & Software PulseEkko Pro system with 100 MHz borehole antennas. The VRP data on borehole four were acquired between 1.2 m and 13.5 m depth, with 0.1 m receiver interval. The ZOP data between boreholes three and four start at 1.25 m depth with 0.25 m vertical interval. For both data sets a 50 ps time sampling interval was adopted. Figure 6 shows the results of the VRP inversions conducted using an MN, an MS and an Occams approach (Minimum Gradient Norm) on the data from borehole four. In all cases, we assumed an error standard deviation equal to 1 ns, consistent with the average difference between two repeated VRP acquisitions on the same borehole, as described, e.g., in Cassiani et al. (2004). The three inversion approaches lead to fairly different moisture content profiles, with Occams inversion leading to the smoothest and MS showing the roughest profile. The comparison against ZOPs results (Figure 6) confirms that the detail in the MS inverted profile is

Figure 5. Borehole GPR data acquired at the Korba experimental site (Tunisia) on June 29, 2010: (a) shows the ZOP data on boreholes three and four, (b) shows the VRP data on borehole four. As borehole three is only about 6 m deep, the ZOP could only be acquired to that maximum depth.

Figure 6. Korba field site. Volumetric soil water content versus depth calculated from MN (gray line), MS (black thick line) and Occam (black thin line) VRP inversions of the data set shown in Figure 5b. The velocity results (dot lines) are converted using the Topp et al. (1980) empirical relationship. The results of the GPR ZOP in Figure 5a are also shown (gray line with squares) for comparison, confirming the pattern detailed by the MS inverted profile.

Downloaded 05 Feb 2012 to 147.162.110.99. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

H16

Vignoli et al. boreholes. In 1994, the site was the scene of an inland crude oil spill following an oil well blowout (Burbery et al., 2004; Christensen et al., 2004; Cassiani et al., 2004) and extensive research was undertaken at the site even since (Kstner and Cassiani, 2009). A thick sequence of poorly sorted silty sands and gravels in extensive lenses, typical of braided river sediments, characterizes the site stratigraphy. An artificial layer of clayey-silty material, less than 1 m thick, originally placed as a liner for rice paddies overlies most of the site. Here, we will analyze only data from borehole BB; the same data we analyzed by Cassiani et al. (2004), where inversion of first arrival times was conducted using Occams approach (Lizarralde and Swift, 1999), and reflections were extracted from the data using a classical approach used in VSP analysis. The GPR acquisitions were performed using a PulseEKKO 100 system with 100 MHz borehole antennas. The transmitter antenna was placed radially with respect to the borehole axis, its center positioned at 0.70 m from the borehole axis; the receiver antenna was lowered at 5 cm steps, starting at 0.75 m below ground. The adopted time sampling interval was 0.4 ns. As described in Cassiani et al. (2004), repeated acquisitions using different transmitter antennas azimuth were used to estimate data errors, under the reasonable assumption of a horizontally isotropic system. An average error of roughly 1 ns was measured as long as the receiver antenna was located above the water table (as deep as 10 m below ground level). When immersed below the water table, the signal repeatability decreased to about 4 ns. Here, we focus on the VRP data acquired on borehole BB on May 6, 2003 (Figure 7). On that date, the water table was at 9.45 m depth below ground. Note how, even in the raw data, some clear reflections are apparent. Figure 8 shows the results of the VRP inversions using an MN, an MS and an Occams approach. Here, similar to the Korba example above, the MS inversion leads to a much more detailed moisture content profile than the MN and Occams inversions. The interesting feature of this data set is that it is possible to compare the VRP inversion results against the location of strong reflections as derived from the extraction of the upgoing field as described in Cassiani et al. (2004). Figure 8b shows the results of the VRP reflection analysis from Cassiani et al. (2004); it is apparent that nearly all the main reflections correspond, as it should be, to strong moisture content (thus, impedance) contrasts as detected by the MS inversion (Figure 8a). Both MN and Occams inversions largely fail to identify such contrasts, so confirming the usefulness of the focused data inversion provided by the MS approach.

not the product of inversion artifacts but is fully consistent with the independently measured strong moisture content variations. Note that in Figure 6 the comparison between ZOP and VRP is overall good, but some discrepancies exist. This is not totally surprising, as ZOP data themselves, if interpreted as done here in terms of direct waves between transmitter and receiver, are subject to artefacts as discussed by several authors (e.g., Rucker and Ferr, 2004; Cassiani and Binley, 2005).

Trecate (Novara, Italy)


In 2003, we acquired high-resolution, time-lapse VRPs at a contaminated site in Trecate, Northern Italy, on a few existing

Figure 7. VRP data acquired on borehole BB at the Trecate site (Novara, Italy) on May 6, 2003.

CONCLUSIONS
In this paper, we present the application of a MS stabilizer to VRP inversion. The performance of the MS stabilizer is compared against MN and Occams inversion approaches, showing that MS is capable of providing velocity profiles with much sharper contrasts. Synthetic and real data sets are analyzed, the latter finding comforting evidence in independent data such as zero-offset profiles and GPR reflections. We can draw the general conclusion that, whenever evidence exists that sharp GPR velocity variations are present, a consistent stabilizer such as MS shall be used in place of more common smoothing stabilizers. In all cases, an accurate estimation of data errors is indispensible to inverting the data using an objective balance between data fitting and a priori model information.

Figure 8. Trecate site. Data collected on borehole BB on May 6, 2003. Comparison between (a) the traveltimes inversion results obtained with different regularizations (MS, MN, Occams), and (b) the double corrected radargram as processed in Cassiani et al. (2004) to extract reflection locations; wt indicates the water table depth.

Downloaded 05 Feb 2012 to 147.162.110.99. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

Focused inversion of VRP traveltimes As future work, it would be interesting to implement and compare the performances and results of other inversion strategies. For example, it would be worth to test on VRP-like data other focused stabilizers like: minimum gradient support (minimizing the support of the spatial derivative of the difference between the solution and the reference model) and total variation. Moreover, as the VSP shares with VRP the same data structure, it would be valuable to test the described MS algorithm on that kind of seismic data.

H17

ACKNOWLEDGMENTS
This research was made possible by funding from the EU Commission (FP7 project MODELPROBE Model driven Soil Probing, Site Assessment and Evaluation, EU project no. 213161 and FP7 project CLIMB Climate Induced Changes on the Hydrology of Mediterranean Basins: Reducing Uncertainty and Quantifying Risk through an Integrated Monitoring and Modeling System, EU project no. 244151.

REFERENCES
Addison, A. D., B. M. Battista, and C. C. Knapp, 2009, Improved hydrogeophysical parameter estimation from empirical mode decomposition processed ground penetrating radar data: Journal of Environmental and Engineering Geophysics, 14, no. 4, 171178, doi: 10.2113/ JEEG14.4.171. Ajo-Franklin, J. B., B. J. Minsley, and T. M. Daley, 2007, Applying compactness constraints to differential travel-time tomography: Geophysics, 72, no. 4, R67R75, doi: 10.1190/1.2742496. Alumbaugh, D., P. Y. Chang, L. I. Paprock, I. R. Brainard, R. J. Glass, and C. A. Rautman, 2002, Estimating moisture contents in the vadose zone using cross-borehole ground penetrating radar: A study of accuracy and repeatability: Water Resources Research, 38, 1309, doi: 10.1029/ 2001WR000754. Annan, A. P., 2005, Gpr methods for hydrogeological studies, in Y. Rubin, and S. S. Hubbard, eds., Hydrogeophysics: Water science technology library service: Springer, 50, 185213. Bertete-Aguirre, H., E. Cherkaev, and M. Oristaglio, 2002, Non-smooth gravity problem with total variation penalization functional: Geophysical Journal International, 149, 499507, doi:10.1046/j.1365-246X.2002 .01664.x. Binley, A., and K. Beven, 2003, Vadose zone flow model uncertainty as conditioned on geophysical data: Ground Water, 41, 119127, doi: 10.1111/gwat.2003.41.issue-2. Binley, A. M., G. Cassiani, R. Middleton, and P. Winship, 2002, Vadose zone flow model parameterisation using cross-borehole radar and resistivity imaging: Journal of Hydrology, 267, 147159, doi: 10.1016/S00221694(02)00146-4. Binley, A. M., and A. Kemna, 2005, DC resistivity and induced polarization methods, in Y. Rubin, and S. S. Hubbard, eds., Hydrogeophysics: Water science technology library service: Springer, 50, 129156. Bniger, U., J. Tronicke, K. Holliger, and A. Becht, 2006, Multi-offset vertical radar profiling for subsurface reflection imaging: Journal of Environmental and Engineering Geophysics, 11, no. 4, 289298, doi: 10.2113/JEEG11.4.289. Brovelli, A., and G. Cassiani, 2010, A combination of the Hashin-Shtrikman bounds aimed at modelling electrical conductivity and permittivity of variably saturated porous media: Geophysical Journal International, 180, 225237, doi: 10.1111/gji.2010.180.issue-1. Burbery, L., G. Cassiani, G. Andreotti, T. Ricchiuto, and K. T. Semple, 2004, Well test and stable isotope analysis for the determination of sulphatereducing activity in a fast aquifer contaminated by hydrocarbons: Environmental Pollution, 129, no. 2, 321330, doi: 10.1016/j.envpol .2003.10.017. Cassiani, G., and A. M. Binley, 2005, Modeling unsaturated flow in a layered formation under quasi-steady state conditions using geophysical data constraints: Advances in Water Resources, 28, no. 5, 467477, doi: 10.1016/j.advwatres.2004.12.007. Cassiani, G., A. M. Binley, and T. P. A. Ferr, 2006, Unsaturated zone processes, in H. Vereecken, et al., eds., Applied Hydrogeophysics: NATO Science Series, Springer, 71, 75116. Cassiani, G., N. Fusi, D. Susanni, and R. Deiana, 2008, Vertical radar profiles for the assessment of landfill capping effectiveness: Near Surface Geophysics, 6, 133142.

Cassiani, G., C. Strobbia, and L. Gallotti, 2004, Vertical radar profiles for the characterisation of the deep vadose zone: Vadose Zone Journal, 3, 10931115, doi: 10.2113/3/4/1093. Christensen, O., G. Cassiani, P. J. Diggle, P. Ribeiro, and G. Andreotti, 2004, Statistical estimation of the relative efficiency of natural attenuation mechanisms in contaminated aquifers: Stochastic Environmental Research and Risk Assessment (SERRA), 18, no. 5, 339350, doi: 10.1007/ s00477-004-0200-x. Clement, W. P., and M. D. Knoll, 2006, Traveltime inversion of vertical radar profiles: Geophysics, 71, no. 3, K66K75, doi: 10.1190/ 1.2194527. Constable, S. C., R. L. Parker, and C. G. Constable, 1987, Occams inversion: A practical algorithm for generating smooth models from electromagnetic sounding data: Geophysics, 52, 289300, doi: 10.1190/ 1.1442303. Davis, J. L., and S. P. Annan, 1989, Ground-penetrating radar for highresolution mapping of soil and rock stratigraphy: Geophysical Prospecting, 37, 531551, doi: 10.1111/gpr.1989.37.issue-5. Deiana, R., G. Cassiani, A. Kemna, A. Villa, V. Bruno, and A. Bagliani, 2007, An experiment of non invasive characterization of the vadose zone via water injection and cross-hole time-lapse geophysical monitoring: Near Surface Geophysics, 5, no. 3, 183194, doi: 10.3997/1873-0604 .2006030. Deiana, R., G. Cassiani, A. Villa, A. Bagliani, and V. Bruno, 2008, Model calibration of a water injection test in the vadose zone of the Po River plain using GPR cross-hole data: Vadose Zone Journal, 7, no. 1, 215226, doi: 10.2136/vzj2006.0137. Farquharson, C. G., 2008, Constructing piecewise-constant models in multidimensional minimum-structure inversions: Geophysics, 73, no. 1, K1K9, doi: 10.1190/1.2816650. Fletcher, R., 1981, Practical methods of optimization: John Wiley & Sons. Giannopoulos, A., 2005, Modelling ground penetrating radar by GprMax: Construction and building materials, 19, no. 10, 755762, doi: 10.1016/ j.conbuildmat.2005.06.007. Greaves, R. J., D. P. Lesmes, J. M. Lee, and M. N. Toksz, 1996, Velocity variations and water content estimated from multi-offset, groundpenetrating radar: Geophysics, 61, 683695, doi: 10.1190/1.1443996. Hadamard, J., 1902, Sur les problmes aux drivs partielles et leur signification physique: Princeton University Bulletin, 13, 4952. Hardage, B. A., 2000, Vertical seismic profiling: Principles: 3rd ed., Pergamon. Hubbard, S. S., J. E. Peterson, Jr., E. L. Majer, P. T. Zawislanski, K. H. Williams, J. Roberts, and F. Wobber, 1997, Estimation of permeable pathways and water content using tomographic radar data: The Leading Edge, 16, 16231628, doi: 10.1190/1.1437539. Jackson, D. D., 1979, The use of a priori data to resolve non-uniqueness in linear inversion: Geophysical Journal Royal Astronomical Society, 57, no. 1, 137157, doi:10.1111/gji.1979.57.issue-1. Kstner, M., and G. Cassiani, 2009, ModelPROBE: Model driven soil probing, site assessment and evaluation: Reviews on Environmental Science and Biotechnology, 8, no. 2, 131136, doi: 10.1007/s11157009-9157-z. Kerrou, J., P. Renard, and J. Tarhouni, 2010, Status of the Korba groundwater resources (Tunisia): Observations and three-dimensional modelling of seawater intrusion: Hydrogeology Journal, 18, 11731190, doi: 10.1007/s10040-010-0573-5. Kouzana, L., A. Ben Mammou, and M. Sfar Felfoul, 2009, Seawater intrusion and associated processes: Case of the Korba aquifer (Cap-Bon, Tunisia): Comptes Rendus Geosciences, 341, no. 1, 2135, doi: 10.1016/j.crte.2008.09.008. Last, B. J., and K. Kubik, 1983, Compact gravity inversion: Geophysics, 48, 713721, doi: 10.1190/1.1441501. Lelievre, P., D. Oldenburg, and N. Williams, 2009, Integrating geologic and geophysical data through advanced constrained inversions: Exploration Geophysics (Collingwood, Australia), 40, 334341, doi: 10.1071/ EG09012. Lizarralde, D., and S. Swift, 1999, Smooth inversion of VSP traveltime data: Geophysics, 64, 659661, doi: 10.1190/1.1444574. Looms, M. C., A. Binley, K. H. Jensen, L. Nielsen, and T. M. Hansen, 2008a, Identifying unsaturated hydraulic parameters using an integrated data fusion approach on cross-borehole geophysical data: Vadose Zone Journal, 7, no. 1, 238248, doi: 10.2136/vzj2007.0087. Looms, M. C., K. H. Jensen, A. Binley, and L. Nielsen, 2008b, Monitoring unsaturated flow and transport using geophysical methods: Vadose Zone Journal, 7, no. 1, 227237, doi: 10.2136/vzj2006.0129. Ludwig, R., A. Soddu, R. Duttmann, N. Baghdadi, S. Benabdallah, R. Deidda, M. Marrocu, G. Strunz, F. Wendland, G. Engin, C. Paniconi, F. Prettenthaler, I. Lajeunesse, S. Afifi, G. Cassiani, A. Bellin, B. Mabrouk, H. Bach, and T. Ammerl, 2010, Climate-induced changes on the hydrology of Mediterranean basins A research concept to reduce uncertainty and quantify risk: Fresenius Environmental Bulletin, 19, 23792384.

Downloaded 05 Feb 2012 to 147.162.110.99. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

H18

Vignoli et al.
Topp, G. C., J. L. Davis, and A. P. Annan, 1980, Electromagnetic determination of soil water content: Measurements in coaxial transmission lines: Water Resources Research, 16, 574582, doi: 10.1029/ WR016i003p00574. Tronicke, J., 2007, The influence of high frequency uncorrelated noise on first-break arrival times and crosshole traveltime tomography: Journal of Environmental & Engineering Geophysics, 12, no. 2, 173184, doi: 10.2113/JEEG12.2.173. Tronicke, J., K. Holliger, W. Barrash, and M. D. Knoll, 2004, Multivariate analysis of cross-hole georadar velocity and attenuation tomograms for aquifer zonation: Water Resources Research, 40, W01519, doi: 10.1029/2003WR002031. Tronicke, J., and M. D. Knoll, 2005, Vertical radar profiling: Influence of survey geometry on first arrival traveltimes and amplitudes: Journal of Applied Geophysics, 57, no. 3, 179191, doi: 10.1016/j.jappgeo.2004 .11.001. Turesson, A., 2006, Water content and porosity estimated from ground-penetrating radar and resistivity: Journal of Applied Geophysics, 58, no. 2, 99111, doi: 10.1016/j.jappgeo.2005.04.004. Vereecken, H., A. Binley, G. Cassiani, I. Kharkhordin, A. Revil, and K. Titov, 2006, Applied hydrogeophysics: Springer-Verlag. Vogel, C. R., 2002, Computational methods for inverse problems: SIAM. Wijewardana, Y. G. N. S., and L. W. Galagedara, 2010, Estimation of spatiotemporal variability of soil water content in agricultural fields with ground penetrating radar: Journal of Hydrology, 391, 2433, doi: 10.1016/ j.jhydrol.2010.06.036. Zhdanov, M. S., 2002, Geophysical inverse theory and regularization problems: Elsevier Science & Technology. Zhdanov, M. S., G. Vignoli, and T. Ueda, 2006, Sharp boundary inversion in crosswell travel-time tomography: Journal of Geophysics and Engineering, 3, 122134, doi: 10.1088/1742-2132/3/2/003.

Moret, G. J. M., W. P. Clement, M. D. Knoll, and W. Barrash, 2004, VSP traveltime inversion: Near surface issues: Geophysics, 69, 345351, doi: 10.1190/1.1707053. Nagarajan, S. S., O. Portniaguine, D. Hwang, C. Johnson, and K. Sekiharac, 2006, Controlled support MEG imaging: NeuroImage, 33, no. 3, 878885, doi: 10.1016/j.neuroimage.2006.07.023. Paniconi, C., I. Khlaifi, G. Lecca, and J. Tarhouni, 2001, A modelling study of seawater intrusion in the Korba coastal plain, Tunisia: Physics and Chemistry of the Earth, (part B), 26, no. 4, 345351, doi: 10.1016/ S1464-1909901)00017-X. Portniaguine, O., and M. S. Zhdanov, 1999, Focusing geophysical inversion images: Geophysics, 64, 87487, doi: 10.1190/1.1444596. Ramos, F. M., H. F. Campos Velho, J. C. Carvalho, and N. J. Ferreira, 1999, Novel approaches on entropic regularization: Inverse Problems, 15, 11391148, doi: 10.1088/0266-5611/15/5/302. Roth, K., R. Schulin, H. Fluhler, and W. Attinger, 1990, Calibration of time domain reflectometry for water content measurement using a composite dielectric approach: Water Resources Research, 26, 22672273, doi: 10.1029/WR026i010p02267. Rubin, Y., and S. S. Hubbard, 2005, Hydrogeophysics: Springer. Rucker, D. F., and T. P. A. Ferr, 2004, Correcting water content measurements errors associated with critically refracted first arrivals on zero offset profiling borehole ground penetrating radar profiles: Vadose Zone Journal, 3, no. 1, 278287, doi: 10.2113/3.1.278. Strobbia, C., and G. Cassiani, 2007, Multilayer GPR guided waves in shallow soil layers for the estimation of soil water content: Geophysics, 72, no. 4, J17J29, doi: 10.1190/1.2716374. Taylor, J. R., 1997, An introduction to error analysis: University Science Books. Tikhonov, A. N., and Y. V. Arsenin, 1977, Solution of ill-posed problems: Winston & Sons.

Downloaded 05 Feb 2012 to 147.162.110.99. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

Вам также может понравиться