an appropriate approach to data acquisition and processing.
With the aim of evaluating the benefits and limits of GPR inspections, Swiss Federal Rail-
Ž .
ways SBB placed an order with EMPA for the investigation of three railway lines. The focus
of the radar inspections was on ballast thickness and on the detection of zones where subsoil
material had penetrated into the ballast. The total length of the inspected sections was 15.1
km. In addition to the 41 trenches based on preliminary GPR results that were available dur-
ing interpretation of radar data, SBB performed their usual investigation programme digging 77
trenches. This provided the opportunity for a quantitative evaluation of radar results.
2. Data acquisition
Data were acquired in summer 97 using a GSSI SIR-10A system and a 900-MHz antenna
Ž .
Ž .
GSSI Model 3101D . Hanninnen et al. 1992
¨
Ž .
pp. 22 provide more information on the 900- MHz antenna. Fig. 1 shows the survey wheel
and the antenna mounted to a trailer which was pulled by a small diesel locomotive at 10 kmrh.
The acquisition parameters can be summa- rized as follows:
Ø acquisition speed: 10 kmrh, Ø horizontal sample rate: 15 scansrm,
Ø data word length: 16 bit, Ø samples per scan: 512,
Ø antenna height: 8 cm, top of sleeper to bot-
tom of antenna casing, Ø antenna orientation: at right angles to travel-
ling direction, Ø scan length: 25 ns.
In the weeks prior to the radar surveys and during data acquisition there was heavy rainfall.
Data were recorded and stored on tape with- out any processing. Also, no effort was made to
avoid data collection when passing over sleep- ers as this would not only require a more so-
phisticated acquisition system but would also
Fig. 1. Set-up for mobile data acquisition on railway tracks.
reduce the potential for future high speed data acquisition.
The site locations are shown in Fig. 2. The different sections are presented in Table 1.
3. Data processing
Raw data were copied to a PC for processing. The aims of this processing can be summarized
as follows. Ø establishment of the ballast surface as a com-
mon reference level, Ø enhancement of signalrnoise ratio,
Ø introduction of coordinate system used by SBB,
Ø reduction of effects caused by sleepers.
Fig. 2. Location of GPR sites.
The following processing steps were applied: 1. bandpass filtering,
2. correction of surface reflection to time zero, 3. migration,
4. horizontal scaling, 5. stack, eight-fold,
6. background removal, 7. horizontal smoothing,
8. vertical gain correction, 9. editing
10. transformation into SBB coordinates. A comparison between raw and processed
data is presented in Fig. 3. For ease of compari- son, only every fifth trace has been plotted in
Table 1 GPR sections
Ž .
Site Name
Length km Trenches available
Total number Remarks
for interpretation of trenches
1 Au–Waedenswil
2.4 4
19 1
Waedenswil–Au 2.4
5 20
2 Rubigen–Guemligen
4.0 12
31 3
Deitingen–Wangen 2.3
8 15
3 Wangen–Deitingen
2.3 6
15 3
Wangen Track 3 0.7
3 8
railway station 3
Wangen Track 4 0.95
3 10
railway station Total
15.1 41
118
Ž .
Ž .
Fig. 3. Comparison between dataset before top and after processing bottom , site 2, profile length 40 m.
the raw dataset and the processed dataset is displayed before transformation into SBB coor-
dinates. Profile length is 40 m. Besides an obvious improvement of the signalrnoise ratio,
the surface reflection, around 3 ns in the raw dataset, was shifted to time zero and eliminated
by applying background removal. The influence of the sleepers, which caused a very irregular
reflection pattern in the raw data, was reduced considerably.
Whereas processing aims such enhancement of the signalrnoise ratio, correction of surface
reflection or coordinate transformation are com- mon to many GPR investigations, the reduction
of the effects caused by sleepers is specific to railway track inspections and is therefore de-
scribed in more detail. This reduction is achieved mainly through processing steps 3–6. In order
to demonstrate the effects of these steps a subset of the dataset presented in Fig. 3 has been
processed for demonstration purposes and is shown in Figs. 4–8. This subset was taken from
the left part of the section presented in Fig. 3 where there are two reflections of interest around
10 ns.
After bandpass filtering and correction of the surface reflection the two reflections are still
Ž .
masked by noise Fig. 4 . Data were migrated
Fig. 4. Dataset after bandpass filtering and correction of surface reflection to time zero, profile length 2 m.
Fig. 5. Dataset after migration.
using a signal velocity of 1.4 = 10
8
mrs. This velocity was chosen after a comparison of radar
data with trench information. As a result the horizontal alignment of the two reflections is
Ž .
improved Fig. 5 and energy has been moved from traces which are covered by sleepers into
traces which are not. The dataset after migration shows increased
near surface
signal amplitudes
for traces
recorded on sleepers. A time gate was defined Ž
. 1.06–5.09 ns and the mean amplitudes within
this gate were scaled to the same level. By doing so, the amplitudes for the time range
below the selected time gate are scaled down
Fig. 6. Dataset after horizontal scaling.
Ž .
Fig. 7. Dataset after stack eight-fold .
Ž .
for sleeper traces Fig. 6 . Please note that this scaling was not applied to all datasets because
not all datasets showed a clear difference in near surface signal amplitudes between sleeper
and non-sleeper traces. Ž
. After stacking eight-fold Fig. 7 , the two
reflections of interest show only minimal reduc- tion in resolution when compared to the dataset
in Fig. 4. As near-surface reflections caused by sleepers show little variation horizontally, back-
Fig. 8. Dataset after background removal.
Fig. 9. Example of poor data quality within railway station, site 3, profile length 200 m.
ground removal is an efficient tool for reducing Ž
. their amplitude see Fig. 8 .
4. Signal velocity