4.3
Discussion and Comparison to Regular GPR Pro-
cessing
In order to judge the utility of theRE results, Fig. 4.12 shows the radar data processed with more standard GPR processing techniques for comparison. More specifically, the processing used for Fig. 4.12 involves three steps: dewow, bandpass filtering (0/5 - 40/60 MHz), and AGC (maximum gain = 2000) (Riger-Kusk, 2011). This kind of processing resolves more small-scale features (e.g. internal layers) than the results with RE (compare with Fig. 4.5), but it also requires more effort, particularly in terms of a priori information, such as the actual bandwidth of the emitted pulse to select appropriate filter parameters, which is not necessarily available until after the data has been acquired. However, Initial analysis with the aim of merely checking whether anything at all or one reflection in particular can be seen in the data (e.g. the reflection from the glacier bed at a certain depth), usually only includes a few basic steps. For the pulse EKKO system these tend to be the dewow correction and a gain function. Typical types of gain functions are the AGC (automatic gain control), which attempts to achieve a constant mean signal level within a fixed time window throughout the data, and the SEC (spreading & exponential compensation), which uses an exponential function to account for geometrical spreading losses and signal attenuation (see also Sect. 2.5.2). Both of these gain functions commonly require the user to identify a maximum gain value and one additional parameter. For the AGC this parameter is the time window width which is typically chosen to equal the pulse width of the system, while the SEC requires the input of an attenuation coefficient for the exponential function (Sensors and Software Inc., 2006).
Selecting appropriate gain function parameters is a crucial point when performing initial analysis of acquired data in the field. As the received signal decays rapidly with time, the (absolute) amplitude values measured tend to span several orders of magnitude, four in the present case. Accordingly, the most useful gain function and gain function parameters (such as maximum gain, decay coefficient, etc.) depend on the details of the system setup such as transmitter power, transmitter-receiver geometry, receiver gain, and the dielectric properties of surface and subsurface materials. The latter are not generally well known prior to a survey and can have a significant impact on the decay of the signal with time/depth, and thereby on the appropriate gain-function parameters. Since the clarity of a reflection depends on its contrast relative to the no-reflection background (volume scattering), i.e. the relative magnitudes of the amplitude gradients after applying a gain, it is not always trivial to identify appropriate gain-function parameters in the field on the spot. Too much gain can over-amplify both the reflection amplitude and the background noise, making an actual reflection hard to discern, while too little gain may lead to the same problem in reverse. The
RE inherently remediates this issue as it is applied before using any kind of gain function and creates relatively high-contrast images since all resulting values lie between zero and one rather than covering several orders of magnitude. The latter is usually the case in raw data, and deep reflections that might be relatively strong (compared to the background of volume scattering) can nevertheless be hard to recognise even after applying a gain. The small range of RE values can make such amplitude gradients easier to identify straight away, since the value of the RE is independent of the absolute magnitude of the gradient in the data. For
Figure 4.12: Radargram processed in a more standard GPR fashion. The processing steps applied are: dewow, bandpass filtering (0/5 - 40/60 MHz), and AGC (maximum gain = 2000). After Riger-Kusk (2011).
example, an increase in the signal amplitude by one order of magnitude – be this from 1000 to 10000 or from 10 to 100 – will result in a similar value of RE (assuming that a variable bin width is used). So all that is required to get a noticeable change in RE is a sufficiently wide (see also Section 4.1) amplitude gradient, independent of the absolute values in that part of the profile. For this reason the RE has the potential to be a useful tool for initial analysis of GPR data.
To make theRE methodology more easily and widely applicable for fast initial analysis of GPR data, it is important to develop guidelines for its usage that depend on the details of the GPR system used for data acquisition, i.e. guidelines for determining the window size parameters. Similar to the parameters of gain functions like the AGC or the SEC, these generally depend on the details of the system setup such as centre frequency, bandwidth, sampling interval, trace interval, transmitter-receiver separation, and the dielectric charac- teristics of the ground. However, while the latter are one of the most important factors when determining the parameters of standard gain functions, the window size for the RE can be determined purely from a priori knowledge of the system setup and can even be generalised to most GPR systems.
Most (commercial) pulsed GPR systems have a pulse width of 2-3 wavelengths at centre frequency and a resolution on the order of one such wavelength(Rial et al., 2009). The sam- pling frequency is usually chosen to ensure an adequate sampling of the centre frequency, i.e. between eight and twelve samples per cycle. Combined with the minimum number of data
4.3. Discussion and Comparison to Regular GPR Processing
Figure 4.13: Comparison of (a) dew o w ed 25 MHz data with basic pro cessing in the form of automatic gain co n trol with a maxim um gain of (b) 300 and (c) 3000 and (d) the results with the R E20 h 10 v using 100 bins and no outlier remo v al.points required for adequate statistics (≈ 150) this knowledge can be used to determine a widely applicable window size forRE calculations: 10 vertical (time dimension) and 20 hori- zontal (along-track dimension) data points and a fixed number of 100 bins. This should give serviceableRE-results for most pulsed GPR systems, independent of their centre frequency. The total of 200 data points per histogram ensures robust statistics of the calculations. Setting a fixed and relatively high number of bins ensures good contrast between gradient regions and the background of volume scattering and improved performance compared to the optimal binning approach. Applied in this manner, the RE can serve as an alternative (almost) parameter-free pseudo gain function for quick analysis of GPR data in the field.
The effect and possible advantage of using the RE20h10v as a gain function is illustrated
by direct comparison with the automatic gain control (AGC). Figure 4.13a shows the 25MHz
profile with only the dewow filter applied. Due to the large amplitude of the initial signal no other features can be made out. For the pulseEKKO PRO system used to acquire this data the default gain for displaying data while it is being recorded is an AGC with a maximum gain of 300 (Sensors and Software Inc., 2006). This results in the glacier bed reflection becoming visible down to about 4800 ns or about 400 m (Fig. 4.13b). Other features, such as two distinct internal layers between 3300 and 4000 m at a depth of approximately 100
m, can now also be seen. In order to view deeper sections of the glacier bed, the maximum gain of the AGC has to be increased. However, this can lead to ‘over-gaining’. For example, if the maximum gain is set to 3000 (Fig. 4.13c), the bottom reflection can be identified down to about 600 m but the gained noise in the data disguises some of the weaker features such as the two internal layers mentioned before. This is where the gain-like effect of the RE is particularly useful. As long as the signal level is not too low, the RE assigns all gradients a relatively similar value, independent of the absolute changes in amplitude involved. In Fig. 4.13d, the bottom reflection is clearly visible down to approximately 600 m (and more faintly for another 100 m) and the two internal reflections in the upper right section of the profile can also be identified. Figure 4.13d is calculated directly from the data shown in Fig. 4.13a, using the ‘standard window size’ (20h10v) and a fixed number of 100 bins. The envelope calculation is simplified to taking the absolute value of each data point which further increases the performance of theREprocessing. Clearly theRE20h10v-results improve
the visibility of the major features in the data, though some of the weaker internal reflections seen in the top 100 m of Fig. 4.13b, cannot be identified in Fig. 4.13d, due to the overall high RE values near the surface. This shows that, for initial analysis of data during or just after acquisition in the field, the RE20h10v approach can have an advantage over standard
gain functions, as it results in a considerable contrast enhancement without requiring any tuning of parameters.
To illustrate the wider applicability of the ’standard window size’, theRE20h10v is applied
to 50 and 500 MHz data acquired on the McMurdo Ice Shelf in the next section.