trace spacing requirements needed for the com- parison of features in 4D. Ideally, the line spac-
ing should be of the order of the trace spacing to allow for optimum visualization of subsur-
face structures in 3D, but in reality, the trace spacing is mostly much smaller than the line
spacing. For a 200-MHz GPR survey with a cross-line dip of 208 and a subsurface velocity
of 0.07 mrns, we need cross-line spacing of 60 cm to avoid spatial aliasing. As for most 4D
surveys, these kinds of parameters are not at- tainable and will degrade the data analysis.
The main complexity of analyzing 4D radar data is of course the processing and visualiza-
tion. These are discussed using the 200-MHz and 500-MHz Borden data sets.
3. The Borden data set
Ž .
The Borden data set collected in 1991 is the most carefully collected and comprehensively
described 4D GPR data set on which data have Ž
been published Brewster and Annan, 1994;
. Sander, 1994; Sander and Olhoeft, 1995 . This
data set, consisting of a closely spaced grid of 2D common offset profiles of 200 MHz and 500
MHz GPR data over a controlled spill was
Ž .
collected at the Canadian Forces Base CFB Borden in July 1991. At this base, an imperme-
able cell was constructed within a saturated sand aquifer with surface dimensions of 9 = 9
m
2
and with a depth of 3.3 m. 770 l of Tetra- Ž
. chloroethylene
TCE , a dense non-aqueous Ž
. Ž phase liquid
DNAPL Pankow and Cherry,
. 1995 , was released in the center of the cell.
The medium in the cell in which the TCE was released is 3.3 m of a medium to fine-grained
beach sand underlain by 3 m of medium stiff clay. A full description of the release experi-
Ž .
ment is given in Brewster and Annan 1994 Ž
. and Brewster et al. 1995 and references therein.
Eight East–West lines and eight North–South lines of GPR data were collected along profiles
on an orthogonal grid with 1 m gridspacing. To characterize the background of the site, data
were collected before the spill and then at 32 Ž
. times after the spill up to 1000 h post-spill .
Initial times between data collection was 8 h, while later in the experiment, the times between
data acquisition increased to about 40 h. For this paper, parts of the 200-MHz and the 500-
MHz data sets were analyzed and both are used to illustrate different processing and visualiza-
tion steps.
4. Data analysis: processing, data differenc- ing and data visualization
While the obvious link between the applica- tions of 4D software in a seismic and a radar
environment facilitates the analysis of 4D radar, there are some complicating differences. These
are due to the differences in size of the data and the number of data sets and data redundancy.
Typically, in seismic, we deal with a fairly
Fig. 2. A typical AVS network for data IrO, crop, differ- ence operation and 3D visualization. The cutout of the
cube of differences can be selected a priori and modified in the ‘‘brick visualizer’’ module.
Ž .
Ž small three to four number of large several
. 100’s–1000 GB surveys. In radar, the cost of
surveys is relatively minimal, so we can have Ž
. tens of relatively small 100’s–1000’s MB sur-
veys, for example, 33 in case of the 200-MHz Borden data set.
Since we can afford to collect multiple data sets with GPR, we can thus combine multiple
different subtraction mechanisms. In effect, in Ž
. the case of n data sets, we have Ý n y 1
differences available. For 3 data sets, this is 3; for 5, it is 10; and for 33, this is 528. Note that
this assumes that we are only looking at simple
Ž differences between two surveys
i.e., first .
derivatives . If we start looking at higher deriva- tives, the amount of difference data sets goes up
accordingly. The interpretation of these 528 dif- ferences can obviously not be done using any
kind of manual interpretation, and we thus need an integrated environment for processing and
interpretation. This implies that we need some kind of both data reduction and automatic anal-
ysis. Even at a fairly modest 20 MB size per survey, this would produce 10 GB of data using
first derivatives. Examining all these data in a profile by profile mode is of course infeasible.
We thus need to pre-process our data with the objective to prepare the data so that each survey
becomes ‘‘similar’’ before differencing and vi- sualizing the data.
Ž .
Fig. 3. Data difference between pre-spill and post-spill 500 MHz data for line E5 at 150 h. Comparison is shown for four Ž . Ž .
Ž . Ž .
different time-shifts. The optimum shift is shown in c . a and d show shifts which are of by 1 in either direction. b shows a shift which is of by 2 from the optimum. Note the spurious energy introduced by a shift of a few samples.
Data pre-processing is done using a combina- tion of in-house and commercial processing and
interpretation packages. One of the core efforts of 4D data acquisition work is to ensure that the
surveys are located at the same place and use the same recording times. Therefore, it is re-
quired to remove any shifts in recording time Ž
. Ž
. depth and space navigational errors among
the surveys. The time differences could have been introduced by changes in instrument pa-
Ž .
rameters different t , coupling changes, in- strument drift, etc. Spatial shifts could have
been introduced by inaccurate navigation. A Ž
first time correlation of the data using first
. arrivals is fairly straightforward. However, it
was found that the exact spatial location for the data is actually a very complex issue due to the
need for high precision. Both the 200-MHz and the 500-MHz Borden
data set, which are used here to illustrate the 4D approach, were carefully controlled and there
was no need for spatial coordinate transforma- tions andror corrections.
For the data differencing and visualization, we use the same package which is being used
by Lamonts 4D seismic group. This commer- cially available software package Advanced Vi-
Ž .
sual System AVS is a UNIX workstation-based visualization programming environment. It en-
ables users to analyze, manipulate and display
Ž . Ž .
Fig. 4. 3D visualization of 200 MHz Borden East profiles. Pre-processed a pre-spill data and b data taken 8 h after spill. Ž .
Ž . Ž .
c Difference between b and a .
large volumes of complex data, including 2D and 3D images, 3D graphics and multidimen-
sional numeric data. AVS utilizes a graphical Ž
. user interface GUI which allows the user to
directly interact with program input and output Ž
. IrO parameters. The user constructs a data
Ž .
analysis network Fig. 2 . The data network
performs not only data IrO, but also parts of the data processing and visualization. As the
connection between modules is taken care of implicitly in this package, research can concen-
trate on processing and visualization instead of code maintenance efforts. The 4D GPR data set
Ž .
of the Borden experiment 33 repeated surveys takes up 46 MB for each of the two profile
directions. Each of these files is loaded into AVS and visualized in a matter of seconds on a
SUN ULTRA30 workstation using a Creator3D graphics card.
The most important function of the 4D-GPR package is the computation of similarities and
differences among regions between the data sets. The differencing process is currently applied as
a simple subtraction operator between the dif- ferent data sets. Future plans include the incor-
poration of region growing in the data sets and subsequent subtraction of the regions between
different data sets. It should be noted that straightforward subtraction is an inherently un-
stable operation due to the wave character of
Ž .
Ž . Fig. 5. 3D visualization of 200 MHz Borden East profiles. Difference between pre-spill data Fig. 4a and data collected a
Ž . Ž .
25 h, b 36 h, and c 45 h after the spill started.
Ž our data i.e., if we are off by 1r2 wavelength,
. our subtraction operator will fail . This of course
is the need behind more complex and stable subtraction operations.
Fig. 3 illustrates the perils of not optimizing the time shift on a 500-MHz profile, by illustrat-
ing the differences between using an optimum Ž
. Ž
shift Fig. 3c and shifts which are 1 Fig. 3a .
Ž .
and d or 2 Fig. 3b from the optimum. For the optimum shift, there is no spurious energy intro-
duced, while for even a shift of a few samples, we introduce ‘‘false’’ differences.
Ž Once an optimum shift is determined which
. can be different for each profile , difference
volumes can be created. Fig. 4 shows how this is done using the 200-MHz data set. All vol-
umes shown in Figs. 4 and 5 have been muted after the first arrival and later than 110 ns,
based on the known area of interest from previ-
Ž .
ous work Brewster and Annan, 1994 . Fig. 4a displays the pre-processed 200 MHz raw data
containing the eight East profiles collected be- fore the spill and Fig. 4b 8 h after the spill
started. Note the strong reflection appearing in an area of the upper North–East corner of the
cut-out in Fig. 4b. The resulting difference be- tween the two displayed data sets is shown in
Fig. 4c, enhancing the differences so that they are much more pronounced. In Fig. 5, three
more difference volumes are shown for 25, 36
Fig. 6. Sketch illustrating the creation of the 3D proxy cube of ‘‘hydrogeological properties’’ from our 4D data set.
and 45 h after the spill began illustrating how Ž
. the spill
strong reflection appears to move
downwards and to spread out in North–South direction along a layer close to the surface. The
3D difference plots create a time-series showing the expansion of the DNAPL throughout the
cell. As a matter of fact, within AVS, it is possible to run through all 32 difference vol-
umes like a movie which illustrates the move- ment of the DNAPL much better than can be
shown in Figs. 4 and 5. Running AVS on the computer further allows cutting out of different
Ž .
Fig. 7. Two cut-outs of the proxy cube for hydrogeological properties 500 MHz .
Ž . Ž .
Ž .
Fig. 8. Time slices at eight different depths a to h through the proxy cube of hydrogeological parameters 500 MHz .
cubes and examining selected plains in arbitrary direction.
5. Creation of a 3D proxy model for hydroge- ological properties