Você está na página 1de 58

Joint inversion of geophysical

data for site characterisation


Deliverable D1.4
Jochen Kamm, Mehrdad Bastani, Laust B. Pedersen,
Daniele Boiero, Corrado Calzoni, Valentina Socco,
Sebastiano Foti, Alberto Godio, Alessandro Arato, Esther
Bloem and Helen French



Deliverable D1.4

Soil Contamination: Advanced integrated characterisation and time-lapse Monitoring

Title Joint inversion of geophysical data for site characterisation
Author Jochen Kamm, Mehrdad Bastani, Laust B. Pedersen, Daniele Boiero,
Corrado Calzoni, Valentina Socco, Sebastiano Foti, Alberto Godio,
Alessandro Arato, Esther Bloem and Helen French
Report No. Deliverable 1.4
I SBN
Organisation name of lead
contractor for this deliverable
UU Uppsala University
No. of pages 48
Due date of deliverable: May, 2011
Actual date of delivery July, 2011
Dissemination level PU
Key words
Joint inversion, modelling, RadioMagnetoTelluric (RMT), Electrical
Resistivity Tomography (ERT)





Title of project: Soil Contamination: Advanced integrated characterisation and time
lapse Monitoring (SoilCAM)

Instrument: 6.3 Environmental technologies, Topic ENV.2007.3.1.2.2, Development of
technologies and tools for soil contamination assessment and site
Contract number: 212663
Start date of project: June 2008, Duration: 48 months


Project co-funded by the European Commission within the Seventh Framework Programme (2007-2013)


Disclaimer
The information provided and the opinions given in this publication are not necessarily those of the authors or the
EC. The authors and publisher assume no liability for any loss resulting from the use of this report.



Deliverable D1.4
Summary

Joint inversion of RadioMagnetoTelluric (RMT) and Electrical Resistivity Tomography (ERT) data were
carried out for the Trecate test site with a view to get an enhanced image of the variability in sedimentary
structure above the ground water table as well the geometry of the underlying aquifer. Whereas ERT data
are very sensitive to changes in the upper few meters the RMT data is more sensitive to deeper
variations, especially variations in the geometry of conductive zones like aquifers. Joint inversion along
profiles of the two datasets thus gives a more complete image than each individual inversion image.
However, it is unclear as to whether the many small details observed in the images for the upper 10
meters can be trusted. Only observations in boreholes can shed light on this question.
In addition to the joint inversion along profiles, for the first time 3D inversion of the RMT data were
attempted. Compared with the 2D models the 3D models show a more consistent and distinct image of
the presumed water table at about 12 meters depth, but less details in the upper few meters than the joint
inversion result. The inversion results were validated by comparing with cross-borehole results at a
single pair of boreholes in the area. A remarkable coincidence between the resistivity in the 3D model
distribution between the boreholes and the 3D image were found in the whole depth interval from the
surface down to a depth of 16 meters. This might indicate that the 3D model is a better image than the
more scattered image obtained from the joint inversion of RMT and VES data, because a 2D model
approach is too simplified even for the relatively undisturbed sedimentary sequence found at Trecate.
Joint inversion of resistivity and radar travel time data have been performed at the Moreppen test site,
bordering Gardermoen airport with a view to get an enhanced image of the time evolution of tracer
concentration in the vadose zone upon snowmelt. The cross-gradient approach was used to search for
models that possibly share a common geometry. The common time-evolution of the models can most
easily be seen when visualising the change of the estimated parameters. The resistivity models show
many small scale anomalies, probably inversion artefacts due to bad data quality. However, on a course
scale a simultaneous and collocated decrease of the respective model parameters can be observed, and
the approximate location of the melt-water front can be located as a function of time in both images. The
strongest dynamic changes occurred over a short time interval from April 3 to April 12 in the depth
interval from 0 to 4 meters.
Due to the conspicuous data quality of the ERT data at the Moreppen site it was concluded that more
advanced attempts to estimate hydrological parameters such as water saturation as a function of time
would not be feasible. Instead synthetic data were constructed to test our new algorithm for estimating
from time lapse measurements both static fields (parameters) such as porosity and dynamic fields
(parameters) such as water saturation. The result was that even with a joint inversion approach and time
lapse measurements it is not possible to determine those fields while at the same time not knowing the
empirical relations like those found in Archies law. The parameters in the empirical laws need to be
estimated from sample measurements in situ or in the laboratory.
In addition to the experiments mentioned above this report also describe additional very high quality
radar measurements conducted between two boreholes at the Trecate site, where also ERT measurements
had been performed. Because of the electrode array used (pole-dipole) for the ERT measurements we
were not able yet to conduct a joint inversion with present version of the program.
Finally a new approach is presented for the joint inversion of seismic P-wave velocities and surface wave
dispersion curves for improved reliability of site characterisation and for estimating the porosity in the
saturated zone in shallow aquifers.






Deliverable D1.4
Content
1. INTRODUCTION ............................................................................................................. 1
2. JOINT INVERSION OF GEOPHYSICAL DATA AT TRECATE SITE ...................... 3
2.1 2D joint inversion of surface ERT and RMT data ....................................................................... 3
2.2 3D individual inversion of RMT data ............................................................................................ 6
2.3 Comparison between 2D joint inversion and 3D individual inversion results ........................... 6
2.4 3D imaging of the aquifer using 3D RMT resistivity model ...................................................... 12
2.5 Discussion on the joint 2D and individual 3D inversion results ................................................ 14
3. JOINT INVERSION OF CROSSHOLE DATA AT MOREPPEN & TRECATE ...... 16
3.1 Implementation .............................................................................................................................. 16
3.2 Moreppen, site and measurement ................................................................................................ 17
3.3 Data acquisition, inspection and processing ............................................................................... 17
3.3.1 Ground penetrating radar (GPR) .............................................................................................. 17
3.3.2 Electrical resistivity tomography (ERT) .................................................................................. 18
3.4 Joint inversion ............................................................................................................................... 21
3.4.1 Joint inversion with Cross-gradient constraints ....................................................................... 21
3.4.1.1 Numerical experiments on synthetic models ......................................................................... 22
3.4.1.2 Results of joint inversion of a field data set .......................................................................... 25
3.4.2. Joint inversion based on physical laws .................................................................................... 32
3.5 Inversion of Cross-hole GPR data in Trecate ............................................................................. 35
4. JOINT INVERSION OF COMPRESSIONAL AND SURFACE WAVES FOR SOIL
POROSITY ESTIMATE ....................................................................................................... 39
4.1Background ..................................................................................................................................... 39
4.2 Inversion Scheme ........................................................................................................................... 40
4.3 Synthetic Example ......................................................................................................................... 41
4.4 Example on real data .................................................................................................................... 43
4.5 Porosity estimate............................................................................................................................ 43
4.6 Conclusions of joint inversion of p-wave and surface wave ...................................................... 44
3. REFERENCES ................................................................................................................... 45

Deliverable D1.4
APPENDIX A JOINT INVERSION RESULTS, MOREPPEN, NORTH WALL OF
LYSIMETER TRENCH ........................................................................................................ 49










1 Deliverable D1.4

1. Introduction
In the frame of the SoilCAM project multiple geophysical datasets have been acquired with
the aim of generating models of corresponding physical parameters e.g. electrical resistivity
or radar velocity to elucidate more details for example about geological structures, geometry
of pollution plume, time variation of contamination front, etc. The geophysical data were
collocated to enable us to either make a joint interpretation or a joint inverse modelling, a so-
called joint inversion. The collected data can be modelled in two- or three-dimensional (2D
and 3D) spaces depending on the data acquisition geometry and parameters. In the last few
decades a vast amount of work has been carried out to develop algorithms for the 2D joint
inversion of multiple geophysical datasets. Vozoff and Jupp (1977) were among the first to
jointly invert two geophysical datasets. They generated synthetic electrical resistivity and
magnetotelluric (MT) datasets from a three-layer model and showed that resolution at certain
depths would be slightly enhanced by including the higher frequency MT data, while by
additional DC data at larger spacing no improvements could be observed. The joint inversion
was then applied to the electrical resistivity and magnetotelluric data collected in a shallow
sedimentary basin. Lines et al. (1988) describe the concepts of cooperative interpretation of
multiple geophysical data by referring to the paper by Golizdra (1980) where three types of
approaches for the inversion of gravity and seismic datasets were explained. The approaches
introduced are separate inversions (no coupling), unified inversions (coupling between
physical properties, e.g. density and seismic velocity) and mixed inversions (assuming a
relationship between the models or structures in the models). Liners et al. (1988) call the latter
joint inversion. Haber and Oldenburg (1997) assume that two models resulting from a joint
inversion should have the same structure and accordingly define a structure operator that
makes use of local spatial gradients or curvature of the estimated model. The operator selected
by Haber and Oldenburg (1997) for the joint inversion is a measure defined as the sum of
squares of differences between the curvatures of the two models in a predefined
domain/space. They then minimize the measure subject to fitting both datasets to a
predetermined level. More recently Gallardo and Meju (2003) introduced the cross-gradient
method to jointly invert Direct Current (DC) electrical resistivity and seismic velocity data in
2D. Later they used the same approach (see Gallardo and Meju, 2007) to make 2D joint
inversion of MT and seismic travel-time data to classify petrophysical and structural
variations based on the resulting electrical resistivity and seismic velocity models. Linde et al.
(2006) used the cross-gradient method combined with a stochastic regularization operator to
invert cross-hole electrical resistivity and ground-penetrating radar data to improve the hydro-
geophysical characterization of Sherwood Sandstone in an area close to Eggborough, North
Yorkshire, UK. Later in another paper Linde et al. (2008) applied the same concept to both
synthetic and field cross-hole GPR and seismic traveltime datasets. They also compare the
results of joint inversions with individual inversions of each dataset by presenting scatter plots
of estimated electrical resistivities and seismic velocities and show that the joint inversion
results are much less scattered than the individual ones.

2 Deliverable D1.4



In this deliverable the results of joint inversion of collocated geophysical data from the
Trecate and Gardermoen test sites are presented. First the resistivity models from the 2D joint
inversion of ERT and RMT data acquired at the Trecate site are shown. The results of
individual 2D inversions of each dataset are presented in Deliverable 1.1. We also show the
resistivity models from the individual 3D inversion of RMT data and compare them with the
2D resistivity models from the 2D joint inversion. Both sets of models are also compared with
the resistivity models from inversion of cross-hole resistivity data prepared by POLITO (see
Deliverables 1.2 and 1.3). Finally the 3D resistivity model is compared to the lithological
information collected in the boreholes located in the study area to make a semi 3D lithological
model over the study area. Then the results from joint inversion of cross-hole resistivity and
IP data carried out by POLITO at Trecate site are presented. Uppsala University and POLITO
conducted cross-hole ERT and GPR data in mid November 2010. Both datasets were jointly
inverted by Uppsala University to study the spatial distribution of smearing zone by analysing
the resulting electrical resistivity and GPR velocity models. Finally, the electrical resistivity
and radar velocity models as a results of the joint inversion of cross-hole time-lapse electrical
resistivity (collected by Bioforsk) and GPR data (acquired by Uppsala University) in
Moreppen area are presented.


3 Deliverable D1.4

2. Joint inversion of geophysical data at
Trecate site
In this section we present the results of joint inversion of collocated geophysical datasets that
were collected with different methods to image details of the geometry of the aquifer as well
as the smearing zone at the Trecate site. Resistivity models from the 2D joint inversion of
surface ERT and RMT data are presented first and then compared to the resistivity models
from individual 3D inversion of the RMT data.

2.1 2D joint inversion of surface ERT and RMT data
As mentioned in the introduction, joint inversion of MT and electrical resistivity data date
back to 1977 (see Vozoff and Jupp, 1977) with the aim to increase the resolution of the
resulting resistivity models by benefiting from the fact that each method has different
resolution and depth penetration. Recently Candansayar and Tezkan (2008) introduced an
algorithm for the joint inversion of ERT and RMT data using smoothness constraints. In their
paper they mention that the RMT data are not very sensitive to the near surface resistivity
variations due to the limited frequency range used, while at the same time ERT data with
small off-sets have relatively high resolution to detect near surface resistivity variations. On
the other hand RMT data carry more information about the deeper structures than the ERT
data. Therefore when the two datasets are jointly inverted more details over a wider depth
range can be extracted from the resistivity models. We have used program EMILIA (Electro-
Magnetic Inversion with Least Intricate Algorithms) for the joint inversion of RMT and ERT
data. The program is developed by Thomas Kalscheuer (see Kalscheuer et al., 2010) in which
the code for the RMT forward and sensitivity calculations is based on the REBOCC program
developed by Siripunvaraporn and Egbert (2000). DCR and RMT data were inverted in 2D
with a smoothness-constrained scheme. Details of the joint inversion can be found in
Kalscheuer et al. (2010). Here we describe some of the parameters used in the joint 2D
inversion of ERT and RMT data.

Location of the ERT and RMT profiles are respectively shown in Figures 2.6 and 2.7 of
Deliverable 1.2. We have made a single figure of the ERT and RMT profile locations (see,
Figure 2.1) to facilitate the referencing made in this deliverable. The ERT and RMT data were
acquired by POLITO and UU, respectively. Profiles E1 and E2 are 108 m long and the ERT
data were collected using 72 electrodes arranged in a Wenner configuration with a minimum
electrode separation of 1.5 m. Profiles E3 to E10 are 96 m long and the ERT data were
collected using the same array configuration and a minimum electrode spacing of 2 m. The
RMT data were collected along the same profiles, marked as R2 to R11, with a station
spacing of 10 m and 11 stations per profile. There were also some extra RMT profiles that
will not be shown in this report.


4 Deliverable D1.4

The results and details of individual inversions of each dataset have been reported in
deliverable D1.1. The joint inversion parameters were adjusted after running a few inversions
and checking the overall datafits including both the ERT and the RMT data. Figure 2.2 is a
schematic that shows the model discretization which is based on the geometry of the field
setup. The lateral cell size within the measuring area is half the minimum electrode separation
(in this case 2 m) used in the ERT measurements to allow for more accurate calculation of the
ERT sensitivity matrix. The cell thickness is increased with depth logarithmically. In the
example shown in Figure 2.2 the rate of increase is by a factor of 10
1/20
meaning that the cells
in row 20 are 10 times thicker than those in row 1. The width of the cells outside the
measuring area (both to the left and to the right) also increases logarithmically.

The regularization used for the inversion is of Occam type where a smooth model is sought
subject to fitting the data to a predefined level. We have used a smoothing anisotropy of 3
where the lateral smoothing is three times the vertical one. The ERT and RMT data were
weighted equally when minimizing the objective function (see Kalscheuer et al., 2010) to
estimate the model parameters, namely resistivity of the cells. The target overall RMS which
is a standard measure of datafit was taken to be 1.

Figure 2.1: Location of ERT (blue text) and RMT (red text) profiles measured in the Trecate site.

5 Deliverable D1.4


Figure 2.2: Schematic of model discritization for the 2D joint inversion of ERT and RMT.

Due to the electrode separation along profiles E1 and E2 (minimum a of 1.5 m) and also the
RMT station spacing it was not possible to carry out a joint inversion of ERT and RMT data
along the two profiles. It can be easily seen in Figure 2.2 that the measuring points should be
collocated with cell nodes. The ERT and RMT setup used do not allow this collocation.
Prior to presenting the results of the 2D joint inversion and individual 3D inversion of the
RMT data we summarize some of the key concepts in RMT data acquisition. As explained in
section 2.1.4 of deliverable D1.1, three components of the magnetic field and two horizontal
components of the electric field from the electromagnetic signal in the frequency range 10-
250 kHz are measured in the RMT method. The two horizontal electric field components and
the vertical magnetic field component are related to the horizontal magnetic field components
via the earth electromagnetic transfer functions, namely impedance tensor (Z) and tipper
vector (T), respectively. In the frequency domain the relationship between is very simple and
is given as

(
(

=
(
(

=
(
(

) ( ) (
) ( ) (
) ( ,
) (
) (
) (
) (
) (
f Z f Z
f Z f Z
f
f H
f H
f
f E
f E
yy yx
xy xx
y
x
y
x
Z Z
(1)
where x, y and f denote the north-south direction, the east-west direction and the signal
frequency, respectively.

6 Deliverable D1.4

Z is a complex tensor and contains information about the electrical resistivity variations in
3D. In the 2D joint inversions we have used the determinant of Z at each frequency to reduce
the 3D effects (see, Pedersen and Engels, 2005).

2.2 3D individual inversion of RMT data
Three-dimensional inversion codes for magnetotelluric (MT) and controlled source MT data
have become more and more powerful, even though they remain computationally expensive.
Recent processing codes have been developed and applied by e.g. Farquharson (2002),
Newman (2003), and Zhdanov (2010) on a variety of targets.
Available for the academic community is a single processor version of the algorithm
WSINV3DMT by Siripunvaraporn 2005, which is a data-space based Occam-like minimum-
structure inversion scheme. Using this code all elements of the impedance tensor shown in (1)
can be included. When applying 3D inversion it is no longer necessary to meet any
dimensionality assumptions. Additionally, any local inhomogeneities can be accounted for in
the modeling/inversion process, whereas in the 2D case only those along the profile direction
could be taken into account, and furthermore the resulting 2D models can be heavily biased
by the distortions caused by off-profile conductivity variations (3D structures).
For convenience, prior to the 3D inversion all the tensor data were rotated by -15 into a local
coordinate system in which the y-axis is along the measured profiles. The model space was
divided into 313128 cells; horizontal cell sizes were designed to allow for two cells
between neighboring sites (5 m long profiles and 4 m across them). The vertical discritization
was the same as in the 2D case (see Figure 2.2), starting with a cell size of 0.5 m. We
included the data of 104 sites on lines R2-R11 (excluding five noisy sites on line 7) and all
nine measured frequencies (10-250 kHz). An error floor of 5% (on off-diagonal elements, Z
xy

and Z
yx
) was applied to the main off-diagonal impedance elements and the same absolute error
floor was applied to the corresponding additional diagonal impedance elements, Z
xx
and Z
yy
,
respectively. The starting model was a homogeneous half space of 1000 Ohm-m. After five
iterations (that each took ca. 8 h on a 2.6 GHz desktop computer) the final model with an
overall RMS of 1.5 was achieved.

2.3 Comparison between 2D joint inversion and 3D individual
inversion results
In order to compare the results from the 2D joint inversion of ERT and RMT data with those
from the individual 3D inversion of RMT data we have chosen the resistivity sections from
the 3D resistivity model where they coincide with the ERT+RMT profiles. Figures 2.3a and
2.3b depict eight resistivity sections extracted from the 3D model along profiles R4-R11 (or

7 Deliverable D1.4

E3-E10, shown in the right columns) and their corresponding resistivity models from the 2D
joint inversion (models in the left column).


8 Deliverable D1.4


Figure 2.3: A comparison between the results of individual 3D inversion of RMT data (models to the
right) and joint 2D inversion of ERT and RMT (models to the left) at Trecate site. a) Models along
profiles R4-R7 (E3-E6) and b) models along profiles R8-R11 (E7-E10).

9 Deliverable D1.4

The resistivity models show similar trends and have clear differences in details. We first
discuss the similarities and then address the differences.

Figure 2.4: Location and Lithological logs of three boreholes B-I, B-J, and B-N located closest to the
area covered by ERT and RMT measurements. The information is extracted from deliverable D3.1.

For the comparison we have also used information from deliverable D3.1 mainly shown in
Figures 2.1-1, 2.1-2a, 2.1-2b, and 2.1-3. We have compiled the information in Figure 2.4 to
compare the resistivity models with the existing information from the boreholes in the area.
Figure 2.5 shows the lithological logs from boreholes B-I, B-J and B-N that were closest to
the study area. According to the information given in deliverables D1.1 and D3.1 at the
Trecate site subsurface consists mainly of low-consolidated sediments (sand, silt and gravel).
Boreholes marked in Figure 2.4 with a B belong to the network that focuses on the plume
core. Stratigraphical descriptions show alternation of gravel in silty sandy matrix and
silty/sandy layers, down to a depth of more than 10 m (deliverable D3.1). Gravel is seen
almost in all boreholes, whereas silty sandy layers are indicated as rather thin layers in the
boreholes. Gravel is found even in the top layer in point B-J, the top layers consist mainly of
sandy silt, sand, silty sand in points B-N, B-O, B-P, with some gravel. The stratigraphic data
in boreholes B-I and B-J show similarly that gravel in a silty, sandy matrix is dominant from

10 Deliverable D1.4

near soil surface down to a depth of more than 6 m. Dominance of gravel layers in sandy silt
matrix results in high permeability in most part of the Trecate site. Top soil and silty sand
layers have lower permeability.
In deliverable D6.1 it is concluded that in the Trecate site there are three main layers that can
be taken into consideration:

- Layer 1: top soil (silty sandy organic composition)
- Layer 2: silty sand layer with gravel
- Layer 3: gravel in silty sandy matrix.
Relatively coarse material (layer 3) is dominant at the site, and layer 2 is not present under the
top soil in some areas.
Both models show a thin low resistivity layer at the topmost part that thins out towards south.
This layer might be interpreted as the siltysandy top soil layer with organic composition (see
Figure 2.4). Models from the 2D joint inversions show more detailed variations compared to
the 3D models and the reason is due to the following facts: a) The ERT data with a 2 m
electrode separation contain more information and have higher resolution power at the surface
than the RMT data and b) the cell width used in the 3D inversion (5 m) is five times coarser
than the width of cells (1 m) in the 2D inversion. This means that the lateral resistivity
variation seen in the 3D model is an average value of 5 cells in the 2D model. A considerably
more resistive layer underlies the first layer that is thicker towards WSW and has a changing
resistivity and thickness. This layer might represent coarser sediments, namely the gravels
shown in Figure 2.4. The resistivity of this layer is generally estimated higher with the 2D
joint inversion than those with the 3D inversion. The second layer appears thicker in the 2D
joint inversion models than the 3D which might be due to the less sensitivity of ERT data to
the lower resistivity below (see synthetic examples shown in Kalscheuer et al., 2010). The
third layer that most probably is the saturated aquifer has a significantly lower resistivity than
the overlying layers. The change in the resistivity might be also related to the changes in the
lithology where, according to lithological logs, below a depth of 10 m more silty sands have
been detected. The third layer is located deeper in the 2D joint inversion models than in
resistivity models from the 3D inversion. In order to validate the estimated depth to the third
layer we have used some of the materials presented in the previous deliverables. In
deliverable D1.3 page 96 the estimated water content from 2D modeling of MRS data
collected by IRIS and POLITO were shown in Figure 6.8. We show the same figure here (see
Figure 2.5) to find qualitative correlations between the water content the resistivity models.

11 Deliverable D1.4


Figure 2.5: 2D distribution of water content (1 = 100%) from 2D inversion of MRS data collected
along a profile in the Trecate site. For the exact location of profile see Figure 5.14 in deliverable D1.3.
This figure is a copy of Figure 6.8 shown in deliverable D1.3.
As seen from Figure 2.5 below a depth of 14-20 m (marked by the first black line) the water
content increases considerably and this depth can be considered as the top of the saturated
zone, named as the first aquifer in deliverable D1.3. The second black line marks the bottom
of high water content zone (Figure 6.8) and below this line the estimated water content drops
clearly. Comparing the water content model with the resistivity models shown in Figure 2.3a
and 2.3b one perceives that the low resistivity zone, namely layer 3, coincides very well with
the highest water content zone, especially in the models extracted from the 3D resistivity
models.
We have also used the results shown in deliverable D1.1 (page 33) where the resistivity model
from the inversion of cross-hole ERT data was compared to the part of resistivity model from
the 2D inversion of surface ERT data. In Figure 2.6 we show these two models together with
the resistivity cross-section extracted from the 3D resistivity model of RMT data along profile
R2 (E1), and also the location of boreholes B-S3 and B-S4 (see Figure 3.4 in D1.1 for more
details). The part of 3D model that coincides with the area within the two boreholes is
enlarged to be compared with the other two cross-sections. As it is clearly seen the resistivity
models from the cross-hole ERT and 3D RMT inversion show more correlation than the
resistivity model from 2D inversion of the surface ERT data. Two distinct features to point at
are the more resistive feature at 4-10 m depth and the underlying lower resistivity zone. It
seems that the depth and resistivity estimates in the cross-hole model and 3D RMT model are
clearly more similar. However, there are obviously some differences between the two models
that are mainly because of the fact that a higher resolution exists in the cross-hole ERT data
and also the coarser block size used in the 3D inversion (partly because of reducing the
calculation time and partly RMT station separation that controls the resolution).

12 Deliverable D1.4


Figure 2.6: A comparison between the resistivity models from the 3D inversion of RMT data, 2D
inversion of cross-hole ERT data and the 2D inversion of surface ERT data in the Trecate site. Models
marked by B and C are taken from Figure 3.3 in deliverable D1.1. Note that the colour scales are the
log10 of the estimated resistivities.

2.4 3D imaging of the aquifer using 3D RMT resistivity model
As discussed above the 3D resistivity model can be used as a reasonably accurate tool to
image the geometry of unsaturated and saturated zones, namely the vadose zone and the
aquifer, in three dimensions. However, effects of regularization used in the inversion process
may always propagate into the final model and therefore this model has to be interpreted with
care. The model, when combined with the other information for example water conductivity,
can provide useful information about variation of porosity in 3D. Since the clay content is
small (see borehole logs shown in Figure 2.4) Archies law provides a good approximation to
construct 3D porosity models from the 3D electrical resistivity model.
The resistivity model from 3D inversion of RMT data is shown in three dimensions in Figure
2.7. The dominating layers mentioned above are clearly seen in this figure. The more
conductive layer at depths of below 10 meters is interpreted to represent the water saturated
layer and the zone above to be the vadose zone. To facilitate the understanding of the
geometry of these layers we have emphasized those cells which lie in a given resistivity
range, which in turn may be related to a change either in lithology or in water saturation.
Figure 2.8a depicts the 3D model where the cells with an electrical resistivity < 300 Om are
clipped. In Figure 2.8b we show two iso-resistivity surfaces of 300 and 1000 Om

13 Deliverable D1.4


Figure 2.7: Resistivity model from 3D inversion of RMT data in the Trecate site.


Figure 2.8: a) Clipped resistivity cell for values below 300 Om. b) Iso-Resistivity surfaces (red 1000
and orange 300 Om) to show the approximate morphology of the unsaturated layer in 3D.

14 Deliverable D1.4


to image the morphology of the unsaturated zone in 3D. Ascan be seen, this zone is thicker in
the central part and deepens towards the south. This is in good agreement with the borehole
information shown in Figure 2.4 indicating that the coarse grain gravel layer is thicker in
boreholes B-J and B-N than in borehole B-I. The morphology of saturated layer is
approximated by making iso-resistivity surface of 100 Om that is shown in Figure 2.9 in three
different views. To see more details the surface with 50 Om is also shown. The thickness of
the water saturated layer reaches its maximum at the centre of the study area.


Figure 2.9: Iso-resistivity surface for 50 and 100 Wm to image the morphology of the saturated
layer/zone in 3D using 3D RMT resistivity model. Three views are presented to show more details.

2.5 Discussion on the joint 2D and individual 3D inversion results
From the results shown in this section we can conclude that
- The joint 2D inversion of ERT and RMT data provided more details of the variation of
electrical resistivity at the surface.
- 3D inversion of the RMT data has a better resolution of deeper conductors and
provides resistivity values that are closer to those measured in the cross-hole
measurements.

15 Deliverable D1.4

- 3D inversion of RMT data is less affected by the near surface inhomogeneities.
In Figure 2.9 we have shown both unsaturated and saturated zones imaged from the 3D
resistivity model and we also refer to the discussion made in Deliverable 1.1 where POLITO
provides some facts about the available data in that area. They write:
The water conductivity is in the range between 450-620 S/cm, with a decrease from the top
of the saturated zone to the deeper level; these values are equivalent to water resistivity of 16-
22 Ohm-m. This leads to an estimate in the saturated zone of the bulk conductivity in the
range of 75-100 S/cm or 100-135 Ohm-m (soil porosity = 0.4; Archie parameters: a=1,
m=2). From page 41 D1.1.
Considering this information and the structure shown in Figure 2.9it is possible to make an
estimate of the porosity variation in 3D to be used for transport modelling.


Figure 2.9: Morphology of saturated and unsaturated zones estimated from 3D RMT resistivity model
in the Trecate area.





16 Deliverable D1.4

3. Joint inversion of Crosshole Data at
Moreppen & Trecate
Time-lapse Crosshole Radar Tomography (GPR) data (by Uppsala University) and Crosshole
Electrical Tomography (ERT) data (by Bioforsk) have been collected. The overlapping period
was the critical phase of a tracer measurement, the snow melt period between 2010-03-29 and
2010-04-21.
We have implemented two different inversion methods for the inversion of time-lapse GPR
and ERT datasets. The first step was to apply the methods to synthetic data (section 3.4.2) to
test their functionality and optimise the algorithm. The first approach is the well-known joint
inversion with cross gradient constraints to enforce structural similarity. It was applied to one
of the field data sets (section 3.4.3).
In the second approach (section 3.4.4), we try to study the feasibility of a joint inversion in
terms of estimation of hydrological subsurface parameters and water saturation state based on
empirical laws with unknown but bounded coefficients, that are determined during the
inversion as well. However, the results were not suitable to apply the method to the field data,
especially when taking the poor data quality of the Moreppen data set into account.
Details of the geological characterisation of the Moreppen test site, design and evaluation of
the tracer experiment, GPR and ERT measurement set-up, GPR data treatment can be found
in Deliverable 1.3, section 4 along with results of the individual interpretations of those data
sets. Only the relevant information will be restated here.
The resistivity and radar velocity models from the individual inversion of cross-hole ERT and
radar data that were acquired in a field campaign by POLITO and UU are delivered.

3.1 Implementation
The implementation of the inversions was carried out in Matlab. The forward routine for ERT
has been developed in FORTRAN by Naser Meqbel in the context of Kalscheuer et al. 2010
with modifications for the borehole configuration following (A. Pidleysecky) and was
interfaced with Matlab as a MEX routine. For travel time data the C-routine PSTOMO_eq6-
62 (Tryggvason 1998) has been interfaced as a MEX routine as well.
The inversion is a general package that allows exchange and combinations of various forward,
transformation, filtering and regularization operators. Using this package, all kinds of joint
and single inversions can be derived as long as the necessary operators have already been
implemented. The core routines of the package currently consist of 11287 code lines in 153
files, not counting the interfaced C and FORTRAN routines. It is fully object-oriented and
therefore easily extendible. The corresponding data processing is tied into this core.

17 Deliverable D1.4

The inversions shown here are all carried out using the Quasi-Gauss-Newton algorithm
(Tarantola 1982). The regularization parameter was adjusted by an automatic OCCAM
scheme (Constable, 1987).

3.2 Moreppen, site and measurement
Data were acquired at Moreppen, near Oslos Gardermoen airport in Norway (Figure 3.1) by
Uppsala University (UU) with assistance from the National Research-Development Institute
for Environmental Protection (ICIM) with an instrument rented from NGI (see Farmani,
2008).
Moreppen is a test site which is geologically similar to the airport site. Hence, in order to
conclude on possible processes at the airport, measurements at Moreppen are valid for
extrapolation.
The geological environment consists of glacial sediments, mostly sands and gravel. The
groundwater table lies at approx. 4 m depth. Two pairs of boreholes (EF and CD) were drilled
through the vadose zone (Figure 3.2) down to the groundwater table. Borehole separation was
2.2 m for the GPR measurements. The borehole pairs for the ERT encompass them with a
separation of 3.2 m. The boreholes are located next to a trench that has been dug before in
order to characterise the site.
On 26.03.2010, before the start of the snow melt period, a highly conductive bromide tracer
was brought upon the snow together with de-icing chemicals (potassium formate and
propylene glycol). Then ERT and GPR measurements were carried out in order to monitor the
transport of the de-icing chemical with the melt water. It is expected that the conductivity
anomaly from the melting water is augmented by the tracer, thus enhancing its visibility in the
resistivity data.
In this chapter, we present results from the processing and inversion of data acquired in the
borehole pair CD (i.e. the South wall of the trench). The results from borehole pair EF (i.e. the
North wall) are shown in appendix A. They are significantly less useful due to worse quality
of the ERT data.

3.3 Data acquisition, inspection and processing

3.3.1 Ground penetrating radar (GPR)

The transmitter and receiver positions were varied to all possible combinations between the
boreholes with a vertical spacing of 20 cm down to 4.2 m depth. During 19 measurement
days, a total number of 9614 radar traces has been collected in borehole pair CD.

18 Deliverable D1.4

We refer the reader to Deliverables D1.1 and D1.2 where we address some of the peculiar
problems with the GPR data inherited and the way we treated them. There, we also describe
the data processing tool and workflow, as it was developed for the GPR data.
A total number of 6619 out of 9614 radar traces could be used. A data error of 0.8 ns was
assumed in order to account for all the reported problems and to minimize the artefacts caused
by previously described systematic errors. The traces recorded with a source or receiver above
1 m depth was mostly contaminated by air waves travelling over instead of through the soil
and could not be used. Below the ground water table (ca.4 m), almost no traces could be picked due
to the strong damping of the water.

Figure 3.1: Moreppen site location (French et al., 2002).

3.3.2 Electrical resistivity tomography (ERT)
Electrodes were installed at vertical positions in the boreholes every 15 cm down to 5 depth.
The recording device used was a Multi-Channel Syscal Pro device by IRIS instruments. The
ERT data were collected with a dipole-dipole electrode configuration and a dipole separation
of 3electrode distances. A total number of 31444 independent (non-reciprocal) apparent

19 Deliverable D1.4

resistivities were measured in the bore hole pair CD. Receiver dipoles were located within the
same borehole as well as the opposite borehole. In order to obtain a reliable estimate of the
data quality, reciprocal measurements have been performed, i. e. transmitter (current
injection) and receiver (potential) dipoles were exchanged. Theory predicts equal values for a
measurement and its reciprocal repetition. Therefore deviations can be considered as an
estimate for the measurement error.
Additionally, the device also reports a data quality factor Q obtained during the stacking
process. This estimate does not account for any systematic (reproducible) error and is usually
orders of magnitude lower than the reciprocal error. Consequently, it gives underestimation of
the data error and is not used in the inversion.
Previously, we reported on the development of a data inspection, visualization, and processing
tool for Crosshole GPR data (see D1.3). A similar tool for the inspection of ERT data has
been developed (Figure 3.3). It allows data visualization in various spatial or spatio-temporal
pseudo section arrangements and interactive or semi-automatic removal of data points (single
data points, electrode related data points or thresholding different quantities like Q, reciprocal
error, apparent resistivity or geometry factor).
The steps of data rejection were: 1) Removal of data points with bad reproducibility Q. 2)
Visual inspection of data. As the errors related to the electrodes may be found due to their
linear signatures in the pseudo sections, the inspection tool was used to spot and remove those
values. Individual outlier values were also removed. 3) Inspection of the reciprocal error
distribution (Figure 3.4). As the distribution is very wide, no Gaussian error model could be
fit to it. As Gaussian statistics, i.e. normally distributed errors, are a prerequisite to our
inversion algorithm, values with reciprocal errors above 10 % were truncated. After that, a
Gaussian distribution could be fit to it. It is clear, though, that the data does in reality not
follow Gaussian statistics. Consequently, even though the fitted distribution has a much lower
standard deviation, we assume a data error of 10% Gaussian noise to account for the
inaccuracy of the error model. Moreover, the 2D assumption of the model may not be valid,
since, in addition to other possible 3D effects, the lysimeter trench is located only ca. 2 m
away from the boreholes. The remaining data set consisted of 9675 apparent resistivities.


20 Deliverable D1.4


Figure 3.2: Borehole positioning in Moreppen in relation to the trench. In boreholes (dots) crosshole
GPR measurements (Uppsala University, inner boreholes, with letters) and crosshole ERT
measurements (Bioforsk, out boreholes) have been carried out.

Figure 3.3: Screenshot of Geoelectrics Visualization and Processing tool.

21 Deliverable D1.4


Figure 3.4: Histograms of relative reciprocal errors. The data is made fit to a Gaussian error model by
a sufficiently high threshold.


3.4 Joint inversion

Geophysical data sets obtained from a common area at the same time are probably inherently
related by the target. Joint inversion is the task of incorporating this similarity assumption into
a combined inversion of several data sets. This might improve the usage of measured data by
reducing the number of models that can explain the individual data sets to those that explain
all data sets while fulfilling a specified assumption of connection between the resulting
models. As the true nature of these connections is generally unknown, the constraints and
their respective levels enforcement have to be appropriate. Otherwise, possible choices can be
structural constraints, such as the Cross-Gradient constraint (Gallardo and Meju, 2007, Linde
et al., 2006) or known or unknown physical relationships between the quantities or empirical
approximations thereof.

3.4.1 Joint inversion with Cross-gradient constraints

Generally, when performing least-squares inversion, a subsurface model is sought that is then
considered to be a particular solution to the inverse problem, when the difference between its
response and the measured data (in a L2-norm sense) falls within the confidence limits of the
measurement. That is equivalent to the condition of achieving an RMS error in the vicinity of
1. Additional constraints may (or have to) be applied to limit the number of possible models,
to stabilise the inversion, i. e. ensuring convergence, and in order to obtain geologically sound
models that match our a priori information.
In the joint inversions of two collocated physical quantities m
1
and m
2
that have been
examined by two individual data sets d
1
and d
2
, we perform a least-squares inversion

22 Deliverable D1.4

simultaneously on both data sets and connect them by applying constraints involving both
models and/or both data sets at the same time.

The cross gradient constraint

t(m
1,
m
2
) =
2 1
m m V V

was described by Gallardo and Meju, 2007. Minimizing its magnitude along with the data
misfits means to minimize cross-product between the gradient fields of the spatially
collocated models along with the data fit. It is optimal (its square norm is zero) when both
model gradients are either parallel or anti-parallel, or when one of the models has no gradient.
Therefore, as soon as one model exhibits a structure in terms of gradients, the other model
will be driven to adapt the same structural features.

3.4.1.1 Numerical experiments on synthetic models

The synthetic experiment is identically set up as the Moreppen experiments, i.e. we used the
same configuration of measurements and boreholes as in Moreppen. The error levels assumed
were 0.8 ns Gaussian noise on the travel-times and 10% error on the apparent resistivities,
which is identical to the assumptions made for the Moreppen data. However, in the synthetic
case, the true error model as well as the true velocity and resistivity model is known (Figure
3.5, right column). Additionally, we omit the time lapse aspect of the data in order to save
inversion run time. We invert the synthetic data sets individually (no cross gradients or other
connection between the data sets or models) and afterwards jointly with cross-gradient
constraints added. Then we compare the resulting models with the synthetic input model.
First, the weighting factors of different constraints were optimised for the single inversions.
For the ERT data, we observed that a strong damping constraint and only a very weak
smoothness constraint (first difference flatness operator) are adequate. These constraints are
described in detail in Menke (1989). We conclude that the ERT sensitivity imposes strong
smoothness by itself. The electric fields are spatially smooth and so are the ERT sensitivities.
On the other hand, the ERT inversion convergences slower than the GPR inversion, which
along with a strong need for damping constraint indicates stronger non-linearity. Additionally,
the sensitivities are concentrated around the electrodes. The data can mostly be explained by
structures around the boreholes. In order to image the volume in between, it was helpful to re-
weight the damping according to the sensitivities, thus punishing changes close to the
electrodes and encouraging them farther away.
The GPR tomography problem behaves very different. It does not need a lot of damping as
the problem is less non-linear. The sensitivities, as they follow from the tracing of rays are not

23 Deliverable D1.4

smooth in model space. Therefore, strong smoothing is necessary to ensure non-scattered
models.
The results of single 2D inversions of each dataset are shown in (Figure 3.5, left column) and
will be discussed later on.
The joint inversion combines the two algorithms as they are, without any changes to the
optimised single inversion algorithms. The only extension is the cross-gradient constraint. The
weight for the cross-gradient term was determined by trial and error.
The results of the joint inversion are shown in (Figure 3.5, middle column).
The thin layer in the central part of the model is highly conductive and of low radar velocity.
Therefore, it is not easy to detect it with both methods. It can be seen that even though both
methods see the layer individually, it is more distinct in the joint inversion image.
Additionally, the geometry of the layer in the individual velocity model is not as accurate as
in the joint inversion model.
The overall joint inversion shows a clearer image of the underlying structure below this
conductive layer, but there are some artefacts introduced that appear in the process of
matching the cross-gradient constraint. For example, at a depth of ca. 1.8 m near the left
boreholes, a high resistivity structure can be seen that is not present in the synthetic model at
all. This structure is more pronounced in the joint inversion model.
The different detailed behaviour of the model can also be analysed more closely by examining
cross-plots of collocated model parameters (Figure 3.6). For a more fair and precise
comparison, only the part of overlapping data coverage is displayed, i.e. between the GPR
boreholes (0 m < x < 2.20 m) and down to ~4m depth. The overall scatter of the points in
parameter space is less for the joint inversion, indicating the systematic connection between
the model parameters as it has been enforced by the cross-gradient constraint. Especially the
part in the centre of the model at intermediate depths is better resolved as the parameter
values obtained come closer to the true values.

24 Deliverable D1.4


Figure 3.5: Results of inversion of synthetic data calculated from the resistivity (top) and velocity
(bottom) models shown in the right column. Left: Results of individual inversions. Centre: Cross-
gradient constrained joint-inversion result. Dots: Electrodes in ERT-boreholes (outer boreholes), GPR
source-receiver locations (inner boreholes).The structure of the synthetic model is overlain with black
and white lines.

Figure 3.6: Cross-plots of resistivity versus velocity for the results shown in Figure 5. Left panel:
individual inversion results. Right panel: Joint inversion. The circles' centres denote the zones
composing of the synthetic input model. Depths are colour-coded.

25 Deliverable D1.4


3.4.1.2 Results of joint inversion of a field data set

Having this high comparability with the Moreppen data enabled us to directly apply the
optimised algorithm to the field data set (see below).
Even though the data set is problematic, a joint inversion could be carried out and the data
could be explained with the error model that we assume to hold for this data set (see section
3.3).

Resistivity models
The resistivity models are displayed in Figure 3.7. They exhibit very high resistivities
between 10
3
and 10
5.5
m. These very high values can only be explained by an extremely dry
or frozen subsurface.
The models can coarsely be partitioned into 4 zones:
1. A very resistive, but possibly not well constrained layer near the surface (<0.5 m,
10
5.5
m,). This layer becomes more conductive at later times ( later ~10
4
m)
2. An extremely resistive vadose zone, that at later times becomes partially very conductive
(0.5 3 m, ~10
5
m, in parts later down to ~10
3.5
m).
3. A less resistive transition zone that shows decrease in resistivity only at very late times.
(3 4 m, ~10
4
m, in parts later down to ~10
3.5
m).
4. The saturated zone (>4 m, ~10
3
m) is still to be considered highly resistive, but more
conductive than the overburden. We observe a slight increase in the resistivity
As expected, the highest resistivities were found above the ground water table at times before
the start of the snow melt. We expect that the water influx as well as the bromide tracer give
rise to lower bulk resistivity. Near the surface, the resistivities start to decrease first. There is a
clear decrease of resistivity in zone 1 between 30.03.10 and 12.04.10. It can also be seen that
the resistivity of zone 2 decreases gradually over course of time, starting at the top and
moving downwards. Before the snow melt, zone 2 is nearly homogeneously resistive but a
horizontally layered structures at ~1.8 m appears after 04.04.10 which becomes gradually
more conductive over time.
Velocity models
The same 4 zones can be identified in the velocity model:
1. A low velocity near surface later (<0.5 m, ~0.1 m/ns, later ~0.08 m/ns).
2. A high velocity vadose zone (0.5 3 m, ~0.13 m/ns, later ~0.1 m/ns).
3. Slightly lower velocity in an intermediate zone (3 4 m, ~0.12 m/ns, later ~0.11 m/ns).
4. Low velocities in the saturated zone (>4 m, ~0.1 m/ns).

26 Deliverable D1.4


The velocity in the zones 1 and 4 is probably not as well constrained as in zone 2 and 3, as
almost no travel times were picked from sources or receivers at those depths. Moreover, low
velocities are generally resolved poorer than high velocities in radar, since the rays travel
according to Fermat's principle, choosing the path that takes least travel time.
As in the resistivity sections, a gradual decrease of velocities can be observed over time.
However, the low velocities enter the model more like a well defined front that propagates
downwards until the high velocity zone is affected almost completely by the decrease of
velocity. The radar velocity decrease is a clear evidence for the change of water content and
thus can be related to water flow and thus used for hydrological modelling.
Between 14.04.10 and 16.04.10 an abrupt increase of the velocity can be observed over all the
model space. It is not clear whether it is a real physical change that has not been recorded as
there is no measurement on the 15.04.10, or whether it is caused by a replacement of the
operators and thus detail changes in the acquisition procedure. As we observe an almost static
shift in the velocities (see D1.3) and as there is no visible change in the resistivity sections of
the corresponding days, the latter is probable argument is more probable.



27 Deliverable D1.4


Figure 3.7: Results of joint inversion with cross-gradient constraints for boreholes C(left) and D(right).
RMS 1.00, RMS
ERT
1.01, RMS
GPR
0.98. The first and third columns are the consecutive resistivity
models, the second and fourth are velocities.

28 Deliverable D1.4


Combined models
As previously outlined, the coarse structure of the models is a similar 3 layer case for
resistivity and velocity. Although some of the detailed structures observed in the resistivities
are not reflected in the much smoother velocity models, there are some details that coincide
very well. For example the structures directly above the ground water table a high resistive
thin layer of higher velocity is located that seems to remain stable through most of the snow
melt, suggesting that it is of lower permeability and thus is circumvented by the water.
Moreover, the thin layer structure at ~3.5 m is present in both models as a conductor and a
layer of slightly lowered velocity, suggesting higher permeability and porosity.
The common time-evolution of the models becomes more obvious when visualising the
change of the estimated values (Figure 3.8). The resistivity models show many small scale
anomalies, which we mostly regard as inversion artefacts. However, on a coarser scale, we
can see a simultaneous and collocated decrease of the respective model parameters in both
models. The interpreted water front is indicated by a horizontal black line in Figure 3.8.
It can also be seen, that the time with strongest dynamics lies between around 03.04.10 and
12.04.10. We learn from this that daily measurements are imperative to resolve this kind of
transport process.

3.4.1.3 Comparison with single inversion results

Even though the data sets have already been inverted separately in a traditional manner (see D
1.3, chapter 4), we repeated the inversions using the algorithm obtained in section 3.4.1.1. in
order to get as comparable images as possible, that is not obscured by a different choice of
constraints or the optimisation algorithm.
The results of the individual inversions are displayed in Figure 3.9. The representation is
identical to the one of the joint inversion results Figure 3.7 to allow for easy comparison of
the detail features.
It is especially obvious that the joint radar image is much more detailed and the transitions
between the zones are sharper. The resistivity models also show slightly more pronounced
structures, but as far as the single inversions contain artefacts, those are more pronounced in
the joint inversions as well.
The correlations of the model parameters can be analysed in cross plots (Figure 3.10) as has
been done in the synthetic example as well. We especially emphasise, that in the joint
inversion the near surface (<0.5 m) is a lot more focussed in parameter space. The same
applies to the zone beneath the ground water table (>4 m). Both areas are not well constrained
by GPR data, so we conclude that the similarity is introduced by the cross gradient constraint.


29 Deliverable D1.4


Figure 3.8: Results of joint inversion, same models and representation as in Figure 3.7, but the change
with regard to the model from 30.03.10 is displayed. Red is a decrease of resistivity or velocity, blue
an increase. Red colours indicate an increase in soil moisture. In every model, a horizontal line is
drawn that outlines the interpreted lower limit of the water plume.

30 Deliverable D1.4


Figure 3.9: Results of individual inversions for comparison with the joint inversion results in Figure
3.7. The first and third columns are the consecutive resistivity models, the second and fourth are
velocities.

31 Deliverable D1.4


Figure 3.10: Cross-plots of resistivity versus radar velocity for selected days from joint inversion
results for borehole pair CD. The depth position of the model parameters is colour-coded.

Figure 3.11: Cross-plots of resistivity versus radar velocity for selected days from individual inversion
results for comparison with joint inversion results in Figure 3.10.

32 Deliverable D1.4

3.4.2. Joint inversion based on physical laws

3.4.2.1 Formulation

By assuming that resistivity and radar velocity can both be described by empirical
relationships to partially common hydrological and petrological parameters, these
relationships can be incorporated into the inverse problem by a model parameter transform.
The relationships considered are Archie's law (Archie, 1942)

bulk
= a
-m
S
water
-n

water
(1)

and the CRIM model (Tinge, 1973)

v
bulkr
= v
soilt
(1 ) +v
water
S
water
+v
air
(1-S
water
) (2)

a, m, n: Archie's coefficients, : porosity, S
water
: water saturation,
water
: water resistivity
(inverse of water conductivity), v
air
, v
water
, v
soil
: radar velocities in pure air, water or soil
material. The model parameters to invert for are

m = [, a, m, n, S
water
,
water
,
soil
]
T


here,
soil
= (v
light
/v
soil
)
2
, and all parameters but
water
are taken to be spatially variable, and
S
water
is considered to vary over time as well. Therefore, we achieve a decomposition of the
problem into a static and dynamic part.
The velocities v
light
= 3x10
8
m/s
,
v
air
v
light
and v
water
v
light
/9

are known constants from the
literature. When formulating the inverse problem in this way, it is straightforward to apply
smoothness constraints to those parameters instead of resistivity and velocity. Moreover,
relatively tight bounds on most of these parameters can be given:
Porosity and water saturation values must lie between 0 and 1.Porosity can often be even
further constrained. For the synthetic experiment, we assume 0.05 < < 0.4. For the soil
permittivity, we assume 3 <
soil
< 20. Bounds on the Archie parameters are given from the
literature: 0.5 < a < 1, 1 < m < 4, 1.5 < n < 2.5. Physically,
water
has to be > 0.
The bounds are enforced by transforming the bounded parameters via a non-linear logarithmic
transform to an unbounded domain. The transform has the form:

m
trans
= log(m lb) - log(ub m) (3)


33 Deliverable D1.4

Inverting for m
trans
instead of m and using the inverse of equation (3) afterwards will always
make sure that lb < m < ub.
For this approach, time-lapse data is required. As we assume all spatially variable parameters
to lie on the same 2D grid with M cells, we get 2M model parameters per day when inverting
for resistivity and velocity, while in this approach only M model parameters (S
water
) for the
each day are added. However, there are 5M (, a, m, n,
soil
) + 1 (
water
) static model
parameters to be considered regardless of the number of measurement days. When using at
least 6 or more days, the number of model parameters in this approach is be smaller than in
the conventional scheme:

2MD > (5+D)M + 1 if D > 5, D measurement days

This indicates that for a longer measurement, the model parameters will be increasingly well
determined. Because of the non-linear nature of the empirical laws a separate resolution of the
model parameters can be expected to a certain degree. However, from a mere repetition of the
measurement on a static water saturation model the non-uniqueness of the model parameters
cannot be resolved. It is necessary that every day independent information is added i.e. that
the water saturation state has to change in order to obtain different states of the empirical
laws. Only then, it may be possible to determine the static variables. Similarly, well
determined static parameters are necessary to obtain a stable estimation of the dynamic water
saturation. Additionally, temporal and spatial smoothness helps to make the problem better
determined.

3.4.2.2 Numerical results

The synthetic input model is displayed in the first row of Figure 3.12 (static parameters) and
Figure 3.13 (temporally variable water saturation). A snow melt scenario similar to the
Moreppen situation has been modelled to examine the capability of the approach. The
synthetic data sets are identically configured as the Moreppen data sets regarding the number
of measurements and the measurement configuration for a particular day. We modelled 15
days with variable water saturation degrees. The synthetic data set consists of 21429 apparent
resistivities and 5165 radar travel times. Note that while the number of radar travel times is in
the same order as it was in the Moreppen experiment, the apparent resistivities considered
useful are approximately two times more numerous. We applied very favourable noise
conditions: 0.3 ns for the travel times, and 5% relative error and 0.01 m absolute error on
the apparent resisitivities. The noise is normal distributed, with the given errors being one
standard deviation. The conditions were chosen favourable in order to explore the limits of the
approach.
The water saturation was chosen to mimic a water plume from the snow melt passing through
the lithology. The model parameters vary on a 2D grid with 28x38 cells. In total, 21281 model
parameters have been used to create the synthetic data and had to be estimated in the inverse
problem.


34 Deliverable D1.4


Figure 3.12: Static parameters. Top: synthetic input model and bottom: recovered model.


Figure 3.13: Water saturation over time. Top: synthetic input model and bottom: recovered
model.

The results are displayed in the lower rows of Figure 3.12 (static parameters) and Figure 3.13
(temporally variable water saturation). The obtained models were obtained after 25 iterations
explain the data to an RMS of 1.14. Later models do fit the data completely (RMS ~1) but
50+ iterations were needed and the models obtained do not visibly differ from the displayed
ones.

Even though the models explain the data down to the very low noise level and tight bounds on
the model parameter ranges, we see that the static Archie parameters and the porosity could
not be very well separated. They are not sufficiently independent. The time-variant water
saturation can be inferred well. The soil permittivity is also clearly resolved, which can be
attributed to the high resolution of the radar data. Figure 3.14 shows the water content
distribution of the synthetic model and the result by calculation of = S
water
for every
water saturation state. We observe that the water content can be recovered robustly, even
though it was not directly inverted for.




35 Deliverable D1.4


3.4.2.3 Conclusions in Moreppen

It could be shown on synthetic examples that joint inversions with structural constraints do
improve structural investigation when it is known that the structures do coincide. By applying
the developed algorithm to a suitable field data set, a reinforced interpretation of the water
transport could be made based on two data sets. Comparisons with the individual inversion
results show the usefulness of joint inversion and validates to a certain extend the structural
similarity assumption for the Moreppen situation.

We conclude that even under very favourable conditions, the parameters necessary to
construct the empirical laws are not determined from GPR and ERT data without adding
additional constraints. These constraints could come from drillings and laboratory
measurements on soil samples.
Since the quality of the field data collected in Moreppen are problematic, we do not consider
testing this approach on them. We also see that an easier model formulated in terms of the
water content could be more promising, as it appears to be a very well determined parameter.



Figure 3.14: Water content over time, derived from results displayed in Figure 3.10 and 3.11.
Top: synthetic input model and bottom: recovered model.


3.5 Inversion of Cross-hole GPR data in Trecate

On 19.10.2010, POLITO and Uppsala collaborated to collect a radar tomography data set at
the Trecate field site. The instrument was provided by POLITO. The two boreholes B-S3 and
B-S4 are separated by 6 m. The source and receiver antennae were moved downwards in their
respective borehole in 25 cm steps between 3 m an 17 m depth. This gives 57 antenna
positions. All possible permutations of source and receiver were measured, resulting in 3249
radar traces. From those 2306 travel times could be picked. The data quality was very good
(see Figure 3.15). The main reason for not being able to pick certain radar traces was the

36 Deliverable D1.4

diminished amplitude with higher offset. The boreholes are relatively close compared to their
depths. The range of the offsets is very large (between 6 m and 15.2 m) and affects the
amplitude of the radar wave strongly. Therefore, large offset traces can be picked with less
precision as short offset traces. We counteract by weighting the obtained travel time by their
respective offset. The error for the i-th picked travel times is assumed to be

i
= Offset
i
/ Offset
min
* 1[ns]

where the Offset
min
= 6 m.


Figure 3.15: Sample of a source gather of radar-gram together with the corresponding picked first-
arrivals (red circles).

The results of the inversion are displayed in Figure 3.16. A clear 4 layer case can be
identified. All the layer boundaries are slightly inclined. The velocity is decreasing
downwards. We interpret:
1. An unconstrained surface layer of intermediate velocity (~0.1 m/ns, down to <3-4 m).
2. A probably dry high velocity zone (~0.11 m/ns, between 3-4 m down to ca. 8 m).
3. An intermediate zone, probable moist due to oscillations in the ground water table or
capillary pressure (~0.085 m/ns, from 8 m to ~11 12 m)
4. Saturated zone (<0.075 m/ns, >11 12 m)

37 Deliverable D1.4


Figure 3.16: Inversion result.


Figure 3.17: A comparison between the results of inversion of a) crosshole ERT and b) cross GPR
data. The zones with low, intermediate and high values are shown in c. Note that the measurements
were not conducted at the same time.


38 Deliverable D1.4

In Figure 3.17 we compare the results of the single inversions of crosshole GPR and cross-
hole ERT (shown in Figure 2.6). Note that the data were not collected at the same time and
the main idea behind the comparison is mainly to show the zones that might partly reflect the
morphologies especially in the vadose zone (depths above 10 m). In the vadose zone we see
three zones marked by LH, LI and HI where the first letter indicates the resistivity and the
second the radar velocity. Letter L, I, and H represent Low, Intermediate and High,
respectively. If we compare this zonation with the lithological log collected in borehole B-I
that is very close to B-S4 the HI zone corresponds to the sandy layer and LH and LI zones
correspond to gravel layer that might partly have silt inside.


39 Deliverable D1.4

4. Joint inversion of compressional and surface
waves for soil porosity estimate

We discuss an approach for joint inversion of seismic velocity (compressional and surface
waves) that can be useful both for improving the reliability of site characterisation and for an
estimate of the porosity in the saturated zone in shallow aquifer.
The approach is based on the data acquisition using the same geophone array and sources of
compressional and surface waves, and then joint inversion of P-wave and surface wave data
sets and finally reconstruction of 2D images of Poisson coefficient distribution and 2D image
of porosity distribution.
We shortly describe the approach by the discussion on a real dataset, that has been recently
acquired in a test site where the methodology has been checked. Were working on new data
set at Trecate site where the suggested methods has to be calibrated with the results given by
other methods, such as interpretation of radar data for estimating the soil moisture (see
previous deliverable).
The joint inversion of seismic data can produce a significant improvement of the result
reliability, mitigating the effect of illness and interpretation ambiguities which are typical of
the two methods. In a joint inversion approach the coupling of different data set can be
performed in several ways accounting for only the geometry or also introducing physical laws
to link unknown parameters (Dal Moro, 2008).
We present here a joint inversion method for refracted P and surface-wave data specifically
designed to deal with laterally varying layered environments which can present strong
velocity contrasts with depth. In point of fact, as seen from other authors (Auken and
Christiansen, 2004), a smooth minimum-structure inversion produces smooth models even for
geological models, which are intrinsically layered. Hence, with respect to other similar
methods (Re et al., 2010), it differs in the parameterization, Frchet derivatives, and inverse
algorithm and it is able to incorporate a-priori information available over the site and any
physical law to link model parameters.
After describing the inversion algorithm we apply the joint inversion to synthetic and field
data sets and finally we discuss the approach for converting P/S wave velocity into porosity.

4.1Background

Seismic surface-wave analysis (SWA) and body-wave refraction tomography (BWT) are both
widely applied to build P- and S-wave (V
P
and V
S
) near surface velocity models, providing
useful information for near-surface geological modelling. SWA is mainly used for V
S
estimation in the field of geotechnics, seismic hazard, vibration propagation and geological
studies (Socco et al., 2010). BWT is widely used to estimate depth and morphology of the

40 Deliverable D1.4

bedrock and to define geological boundaries and property distribution in sedimentary
environments (Zhang and Toksoz, 1998). SWA and BWT present several synergies. The data
can be jointly gathered with a significant optimization of fieldwork, since the acquisition
schemes are very often similar. Furthermore, since they have different depths of investigation,
different intrinsic limitations and different resolution and sensitivity with depth, the coupling
of the two methods helps in building consistent V
P
and V
S
models (Piatti and Socco, 2010).
Moreover the gathered analysis of VP and VS dataset leads to the imaging of mechanical
parameter (e.g. Poisson coefficient) or in sedimentary environment the results can be
converted into distribution of soil porosity
SWA inverse problem is ill-posed, mix determined and strongly non-linear and suffers from a
solution non uniqueness. Thus, different models may supply equally good fitting with
reference to the experimental dispersion curve (Socco and Boiero, 2008). BWT is an ill-posed
and ill-conditioned inverse problem as well. The presence of low velocity layers embedded in
stiffer layers or gradual velocity increases over interfaces with strong velocity contrast
produce hidden layers and significant errors in the final velocity profile regardless of the
algorithms that are used to invert the traveltimes (Palmer, 2010).

4.2 Inversion Scheme
To couple SWA and BWT data we discretise Earths surface (x,y) using a (regular) grid of K
points, and at each point k we consider a layered 1D model m
k
characterised by V
P
(z), V
S
(z)
and, the density (z). In the following we illustrate the scheme for the 2D case but the
extension to 3D case is straightforward. We use a (damped) weighted least square algorithm
to jointly invert dispersion curves and first breaks. This involves minimization of the misfit
function Q:

( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( ) ( )
-1 -1
-1 -1
- - - -
- - - -
T T
obs obs obs prior prior prior
T T
R pr
Q fw fw
pr pr
( (
= + +
( (

( (
+ +
(

d m C d m m m C m m
Rm C Rm pr m C pr m
(1)
where the forward response fw(m) evaluates for each model m
k
the correspondent dispersion
curve using the Haskell (1953) and Thomson (1950) method and the P-wave first breaks
through the finite difference solution of the eikonal equation as proposed by Podvin and
Lecomte (1991), which can compute traveltimes in very contrasted velocity models as a
layered one could be. For P-wave traveltime computation a 2D/3D grid mesh is built by
nearest-neighbour interpolation of all the m
k
. The covariance C
obs
(inferred from the
measurements) weights the observed data d
obs
. The vector m
prior
represents the a-priori
information with the associated covariance C
prior
whereas pr represents the expected value for
a physical property pr(m) of the model parameters (i.e. the Poissons ratio which links V
P
and
V
S
) with its covariance C
pr
. In eq. (1), the regularization consists in the inclusion of R which
helps to minimize the differences between model m
k
and the models at surrounding points. Its
effectiveness depends on the covariance matrix C
R
(Auken and Christiansen, 2004). To
minimise Q, the model solution at the nth iteration can be expressed as

41 Deliverable D1.4


( ) ( ) ( ) ( ) ( ) ( )
-1
-1 -1 -1 -1
1
-1 -1 -1 -1
- - - -
T T T
obs prior R pr obs pr
n n
T T T
obs obs n prior prior n R n pr pr n
fw pr

+
| |
( + + + +

|
= +
|
( | + + +
\ .
G C G C R C R G C G I
m m
G C d m C m m R C Rm G C pr m
(2)
where the Jacobian G represents the sensitivity matrix of the dispersion curves and the first
breaks respect to the model parameters (V
P
(z), V
S
(z)) and where is the damping parameter,
which is used to stabilize the solution. It is worth noticing that the sensitivity of the first
breaks respect to V
P
(z) is evaluated without performing ray-tracing, following the approach
proposed by Ammon and Vidale (1993). This presents several advantages, for instance, the
inversion is less biased by the reference model and could consider finite frequency effects.
Because this method explicitly constructs the sensitivity matrix (Frchet derivatives) by
repeating the forward calculation for each perturbed parameter, this approach is limited to a
small number of model parameters as in the case of a quasi-layered model. The matrix G
pr

contains the partial derivatives of the considered physical property respect to model
parameters.

4.3 Synthetic Example
The synthetic datasets, representing a multi-fold seismic survey, has been generated through a
FEM code (Comsol Multiphysics, stress-strain module), using an axial symmetric scheme
and a Ricker source type (dominant frequency 10 Hz). The model is linear elastic isotropic
(its seismic properties are reported in Table 4.1) and presents topographic unevenness and a
dipping half-space (Figure 4.1a).


Figure 4.1: The synthetic model: a) the model geometry; the sensor positions (red triangles) and the
centres of the window used to estimate the dispersion curves (blue points) are shown; b) the estimated
dispersion curves; c) the picked first breaks and the source positions (black stars).

42 Deliverable D1.4

The finite difference solution of the eikonal equation is computed using FDTIMES by Pascal
Podvin (Gophysique - Centre de Gosciences - ARMINES/cole des Mines de Paris).
Concerning the SWA processing, all seismograms have been windowed in the spatial domain
using a sliding window according to the method proposed by Socco et al. (2009). The f-k
spectra from different shots with the same spatial windowing have been stacked, and the
picking of maxima has been performed on each stacked f-k spectrum, yielding a set of 12
dispersion curves (Figure 4.1b) evenly-spaced (25m) along the survey line (Figure 4.1a). First
breaks for each channel and for different sources have been also picked (Figure 4.1c).
Dispersion curves and first breaks are inverted together using the scheme described by the eq.
2. The adopted reference model is the layered 1D V
P
(z) and V
S
(z) profile described in Table
4.2.

Table 4.1 The seismic properties of the
Model
Layer
V
P

(m/s)
V
S

(m/s)
(kg/m
3
)
1 240 120 1800
2 340 170 2100
3 500 270 2400
Table 4.2 The seismic properties of the Reference
Model
Layer
Thk
(m/s)
V
P

(m/s)
V
S

(m/s)
(kg/m
3
)
1 10 200 100 1800
2 10 400 200 2100
3 - 600 300 2400

The results of the inversion are shown in Figure 4.2, where each sounding corresponds to an
m
k
. The maximum differences with respect the true model are about 10% for thickness and
7.5% for the velocity values (for the second layer which is hidden for P-wave BWT for the
majority of the profiles).


Figure 4.2: The inversion result: a) the estimated S-wave velocity model; b) the estimated P-wave
velocity model. The true model geometry is superimposed.


43 Deliverable D1.4

4.4 Example on real data
A dataset has been acquired with a receiver array made up of 48 vertical 4.5 Hz geophones
with 1.5 m spacing. The acquisition parameters in time domain were 2 s acquisition length
and 0.5 ms sampling rate (Figure 4.3a). Ten (10) evenly spaced dispersion curves (Figure
4.3b) have been extracted using the same scheme adopted for synthetic data using 10 different
Gaussian windows according to the method proposed by Bergamo et al. (2010). The P-wave
first breaks have been visually picked for all the seismograms (Figure 4.3c). The VS and VP
models retrieved through the inversion are shown in Figure 3d and 3e, respectively. Both of
them point out a shallower interface at 2-4 m depth gently sloping to the left and a deeper
interface at 8-12 m depth leaning to the right. The results are in agreement with previous
studies.

Figure 4.3 Field case: a) an example of raw data; b) the dispersion curves estimated along the line; c)
the P-wave first breaks; d) the estimated S-wave velocity model; e) the estimated P-wave velocity
model.

4.5 Porosity estimate
Several approaches are available for estimating the soil porosity from seismic waves dataset; a
exhaustive review on this topic is given by Berryman (1995). We suggest using the approach
given by Foti and al. (2002) based on the development of Biots theory. The approach
assumes the following:
1. in the low frequency range no motion between the fluid and the solid phase is
observed:
2. the soil grains are incompressible.



44 Deliverable D1.4

The soil porosity can be estimated according to the following relationship (Foti et al. 2002):

where n is the value of soil porosity,
s
is the soil density,
f
is the fluid density, K
f
is the bulk
modulus of the fluid, v
sk
is the Poisson coefficient of the solid skeleton, varying between 0.1
and 0.4; it can be shown that the Poisson coefficient of the solid skeleton has a negligible
influence on the final estimate (Foti et al., 2002).

4.6 Conclusions of joint inversion of p-wave and surface wave
The proposed algorithm of joint inversion of surface-wave dispersion curves and P-wave first
breaks has proved to be effective in the case of laterally varying layered sites (models) to
build V
P
and V
S
models. It presents advantages with respect to individual SWA and BWT
since it imposes internal consistency for all the model parameters reducing the required a-
priori assumptions (i.e. Poissons ratio in SWA) and the illness of the two methods (i.e.
hidden layers for BWT).
Were also exploring the reliability and sensitivity of the approach in assessing the soil
porosity in the saturated zone at Trecate site; particularly were dealing with the sensitivity of
the method in improving the data interpretation of georadar data in detecting the gradual
changes of the soil porosity in the most contaminated area as due to the combination of
several effects such as mineralisation and mineral precipitation, pore clogging.














45 Deliverable D1.4

3. References

Ammon C. J. and Vidale J. E. 1993. Tomography without rays. Bulletin of Seismological Society of
America 83, 509-528.

Archie, G. E., 1942.The electrical resistivity log as an aid in determining some reservoir
characteristics. T. Am. Inst. Mineral. Metall. Petrol. Eng., 146, 54-62.
Auken E. and Christiansen A. V. 2004. Layered and laterally constrained 2D inversion of resistivity
data. Geophysics 69, 752-761.

Bergamo P., Boiero D. and Socco L.V. 2010. Retrieving 2D Structures from Surface Wave Data by
Means of a Space-varying Spatial Windowing. 16th EAGE Near Surface 2010. Zurich,
Switzerland. Extended Abstracts.

Berryman, J. G. 1995. Mixture theories for rock properties. In Rock physics and phase relations: A
handbook of physical constants (ed. T. J. Ahrens), pp. 000000. Washington: American
Geophysical Union.

Boiero D., Calzoni C. Socco L.V. 2011, Joint Inversion of Surface-wave Dispersion and Pwave
Refraction Data for Laterally Varying Layered Models, Proc. 73
rd
EAGE Conference & Exhibition
2011, Vienne, Austria.

Candansayar, M.E. & Tezkan, B., 2008. Two-dimensional joint inversion of radiomagnetotelluric and
direct current resistivity data, Geophys. Prospect., 56(5), 737749.

Constable, S. C., Parker, R. L. and Constable, C. G., 1987. Occam's inversion: A practical algorithm
for generating smooth models from electromagnetic sounding data. Geophysics, 52(3), 289-300.
Dal Moro G. 2008. VS and VP vertical profiling via joint inversion of Rayleigh waves and refraction
travel times by means of bi-objective evolutionary algorithm. Journal of Applied Geophysics 66,
15-24.

Farmani, M. B., 2008 A. Estimation of unsaturated flow parameters by inverse modeling and GPR
tomography, PhD thesis.

Farquharson, C. G., D.W. Oldenburg, E. Haber, and R. Shekhtman, 2002. An algorithm for the three-
dimensional inversion of magnetotelluric data, 72
nd
Annual International Meeting of the Society of
Exploration Geophysicists.

Foti S., Lai C.G., Lancellotta R. (2002) Porosity of Fluid-Saturated Porous Media from Measured
Seismic Wave Velocities, Geotechnique, vol. 52 (5), 359-373



46 Deliverable D1.4

French, H.K., C. Hardbattle, A. Binley, P. Winship and L. Jakobsen, 2002, Monitoring snowmelt
induced unsaturated flow and transport using electrical resistivity tomography, J. Hydrol., 267(3 -
4), 273-284.

Gallardo, L. A., and M. A. Meju, 2003, Characterization of heterogeneous near-surface materials by
joint 2D inversion of dc resistivity and seismic data: Geophysical Research Letters, 30, 1658, doi:
10.1029/2003GL017370.

Gallardo, L. A., and M. A. Meju, 2007, Joint two-dimensional cross-gradient imaging of
magnetotelluric and seismic traveltime data for structural and lithological classification:
Geophysical Journal International, 169, 12611272, doi:10.1111/j.1365246X.2007.033-66.x.

Golizdra, G. Y., 1980, Statement of the problem of comprehensive interpretation of gravity fields and
seismic observations: Izvest. Earth Phys., 16, 535-539.

Haskell N. A. 1953. The dispersion of surface waves on multilayered media. Bulletin of Seismological
Society of America 43, 17-34.

Kalscheuer, T., M. Garca, N. Meqbel, and L. B. Pedersen. Non-linear model error and resolution
properties from two-dimensional single and joint inversions of direct current resistivity and
radiomagnetotelluric data. Geophys. J. Int., 182(3):11741188, 2010.

Linde, N., A. Binley, A. Tryggvason, L. B. Pedersen, and A. Revil, 2006a, Improved
hydrogeophysical characterization using joint inversion of crosshole electrical resistance and
ground penetrating radar traveltime data: Water Resources Research, 42, W12404, doi:
10.1029/2006WR005131.

Linde, N., A. Tryggvason, J. Peterson, and S. Hubbard, Joint inversion of crosshole radar and seismic
traveltimes, Geophysics, 73, G29-G37.

Lines, L. R., Schultz, A. K. & Treitel, S., 1988. Cooperative inversion of geophysical-data,
Geophysics, 53, 820.

L.B. Pedersen and M. Engels, 2005. Routine 2D inversion of magnetotelluric data using the
determinant of the impedance tensor: Geophysics, 70, G33G41.

Menke, W., 1989. Geophysical Data Analysis: Discrete Inverse Theory, Revised Edition (textbook),
Academic Press, Inc., New York.

Newman G. A., S. Recher, B. Tezkan, and F. M. Neubauer, 2003. 3D inversion of a scalar radio
magnetotelluric field data set, Geophysics 68 (3), pp 791-802, doi: 10.1190/1.1581032

Palmer D. 2010. Non-uniqueness with refraction inversion a syncline model study. Geophysical
Prospecting 58, 203218.


47 Deliverable D1.4

Piatti C. and Socco L.V. 2010 Joint Inversion of P-waves refraction travel times and surface wave
dispersion curves. 72nd EAGE Conference and Exhibition. Barcelona, Spain. Extended Abstracts.

Pidlisecky, A., and Knight, R., 2008. A Rapid 2.5 Forward Modeling Algorithm for Electrical
Resisitvity Modeling. Computers and Geosciences, 34, 16451654.

Podvin P. and Lecomte I. 1991. Finite difference computation of traveltimes in very contrasted
velocity models: a massively parallel approach and its associated tools. Geophysical Journal
International 105, 271-284.

Re S., Strobbia C., De Stefano M. and Virgilio M. 2010. Simultaneous joint inversion of refracted and
surface waves. 80th SEG Conference and Exhibition. Denver, USA. Extended Abstracts.


Siripunvaraporn W. and G. Egbert, 2000. An efficient data-subspace inversion method for 2-D
magnetotelluric data. Geophysics, 65(3),791803.

Siripunvaraporn, W., G. Egbert, Y. Lenbury, and M. Uyeshima, 2005. Three-dimensional
magnetotelluric inversion: data-space method, Physics of the Earth and Planetary Interiors, Vol.
150(1-3, pp 3-14, doi: 10.1016/j.pepi.2004.08.023

Socco L.V. and Boiero D. 2008. Improved Monte Carlo inversion of surface wave data. Geophysical
Prospecting 56, 357-371.

Socco L.V., Boiero D., Foti S. and Wisn R. 2009. Laterally constrained inversion of ground roll from
seismic reflection records. Geophysics 74, 6, G35-G45.

Socco L.V., Foti S., and Boiero D. 2010. Surface wave analysis for building near surface velocity
models: established approaches and new perspectives. Geophysics 75, 5, A83-A102.

Tarantola, A., and Valette, B., 1982. Generalized nonlinear inverse problems solved using the the least
squares criterion. Reviews of Geophysics and Space Physics, 20(2), 219-232.
Thomson W. T. 1950. Transmission of elastic waves through a stratified solid medium. Journal of
Applied Physics 21, 89-93.

Tinge, W. R., Voss, W. A. G. and Blossey, D. F., 1973. Generalized approach to multiphase mixing
theory. J. Appl. Phys., 44(9), 3897-3902.
Tryggvason, A., 1998. Seismic tomography: inversion for P- and S-wave velocities. Acta Univ. Ups.
Comprehensive Summaries of Uppsala Dissertations from the Faculty of Science and Technology,
369. 31pp. Uppsala. ISBN 91-554-4224-0.
Vozoff, K., and Jupp, D. L. B., 1975, Joint inversion of geophysical data: Geophys. J. Roy. Astr. Soc.,
42, 977-991.

Zhang J. and Toksoz M. N. 1998. Nonlinear refraction traveltime tomography. Geophysics 63, 5,
1726-1737.



48 Deliverable D1.4

Zhdanov,M. S., A. Green, A. Gribenko, and M. Cuma, 2010. Large-Scale Three-Dimensional
Inversion of EarthScope MT Data using the Integral Equation Method, Physics of the Solid Earth,
46, pp 70-678.















49 Deliverable D1.4

Appendix A Joint inversion results,
Moreppen, North wall of Lysimeter trench

Thirteen days of ERT measurements were used and 18 days of GPR measurements. In total 16668 out
of 29198 apparent resistivities were selected for the inversion process after a similar treatment as
described in 3.3.2. Even though the threshold for truncation of reciprocal errors was 10%, a data error
of 15% had to be assumed in order to fit the data within the assumed error model. We could pick
reliable first arrivals from 6672 out of 9936 radar traces. We assumed a data error of 1 ns.
Figure A.1 shows the joint inversion results, the temporal change with respect to the chosen reference
model from 01.04.2010 is displayed in Figure A.2. Cross-plots between resistivity and velocity are
shown in Figure A.3.


50 Deliverable D1.4


Figure A.1: Results of joint inversion with cross-gradient constraints for boreholes E(left) and F(right).
RMS 1.00, RMS
ERT
1.01, RMS
GPR
0.98. The first and third columns are the consecutive resistivity models,
the second and fourth are velocities. The data errors for the ERT were estimated to be 15% and for GPR 1
ns.

51 Deliverable D1.4


Figure A.2: Results of joint inversion, same models and representation as in Figure A.1, but the change
with regard to the model from 01.04.10 is displayed. Red is a decrease of resistivity or velocity, blue an
increase. Red colours indicate an increase in soil moisture. In every model, a horizontal line is drawn that
outlines the interpreted lower limit of the water plume.

52 Deliverable D1.4


Figure A.3: Cross-plots of resistivity versus radar velocity for selected days from joint inversion results for
borehole pair EF. The depth position of the model parameters is colour-coded.

Você também pode gostar