Introduction
Several people have reported an instability in the ravn algorithm which manifests itself as a record to record fluctuation of the polynomial order derived by the program for the robust fit to the gain surface. Seldom is a reason for such variation in fit evident.The above behavior is most noticeable in data recorded over low velocity media [Gulf of Mexico, offshore Nigeria etc.]. These data display a pronounced triangular onset of signal where the initial energy at the far offset does not arrive until more than halfway through the record. The ravn algorithm decimates the observed gain surface prior to the derivation of the robust fit resulting in too few samples from this zone being incorporated into process. In addition there is normally quite an expressed kink in the gain surface within this zone which when combined with the undersampling causes the ravn algorithm to class such data as noise to be ignored in the surface fit. To convince ravn that the affected data is indeed signal a significantly higher number of samples must be gleaned from this zone. This may be accomplished using the -top parameter which allows the user to specify a percentage of the record from which to use all samples in the calculation of the surface fit. Such oversampling allows the algorithm access to the higher amplitude information in affected zone but does not address its' pathologic triangular shape.
Recent work has advanced our understanding of the ravn algorithm, the effects of data preconditioning and the methodology needed to properly parameterize such datasets. This note will update the user to recent changes that have been made to the program as well as demonstrate an systematic parameterization procedure on a real dataset.
The Previous Algorithm
The principal algorithm consists of the following logic:
-
1. compute, using a median agc operator, a gain surface using the input record. The agc window length is the command line entry -w [default=250ms] and the agc reference amplitude is -amp [default = 15% 0f 2047].
2. decimate the gain surface as specified by the command line entries -tstep [default = ntrc/12], -tstart[default =1], -tend [default=last trace], -sstep[default = 100ms] and -top [default = 20%].
3. calculate the polynomial coefficients describing the fit of order N, where N varies from 1 to the limit set by the command line parameter -ord [default = 0] to the agc gain surface. The default setting allows the program to determine when it has tested enough orders to have found the most statistically significant fit. The type of fit, least squares or robust, is controlled by the command line entry -f [default=robust] while the number of iterations used to compute the fit at each order is given by the command line entry -ilim [default=3]
4. calculate a regression coefficient and F-test value to determine statistical significance of the surface fit at this order.
5. repeat steps 3 and 4 over a program determined range of orders, or up to the limit set by -ord above.
6. determine which order tested provides the optimum, statistically significant fit with respect to the input data.
7. perform median trace adjustment if requested by the command line parameter -nomedian [default=used]. The number of iterations of median trace adjustment is given by the command line parameter -iter [default=3].
8. record the optimum surface order and coefficients along with any median adjustment coefficients in the attached coefficients file.
Upgrades
Several changes have been made to the command line parameters. The rational for these changes is covered in the body of the note. They are catalogued here for ease of reference:
-
1. The -nomedian flag has been removed.
2. -median must now be included on the command line should you require median trace adjustment.
3. The default number of iterations for median trace adjustment, if requested, is now 0 [zero]. This means that if you flag ravn with -median you must also specify how many iterations of median trace adjustment you desire using the command line entry -iter.
4. The -tstep default has been set to 1[i.e.every trace].
5. The -sstep default is now 24 milliseconds.
6. A new command line entry, -sord [default=0] has been created which if specified will constrain ravn to fit only surfaces of the specified order for all records on the line. The default is still to have the program determine the most statistically significant order of fit.
Data Preconditioning
As discussed in the introduction above, the surface fit provided by ravn becomes unstable in the face of a deeply muted onset of data. This instability is exacerbated by a large amplitude decay throughout this zone.Recent experimentation with several such datasets has yielded the observation that the stability of the surface fit can be improved if the data is preconditioned prior to running ravn. The following processing steps ought to be considered:
-
1. attenuate direct arrival energy.
2. apply a normal moveout correction
3. apply a tn scalar [n = 1.0]
4. interpolate amplitudes into dead traces.
5. apply an onset mute combined with a low cut filter to remove excessive normal moveout stretch on the far offsets.
Ravn Parameterization
Parameter testing is mandatory. Without it you will have no idea how well the surface calculated by ravn actually fits, in a robust sense, the decay surface of your data. It is likely that without adequate testing the incorrect order of surface will be used resulting in an offset dependent skew being introduced to the data amplitudes. The following testing regimen has been found reliable:
-
1. edit a subset of records from your data for testing such that you maintain 100% [single fold] coverage at the zone of interest. The record interval is of course a function of acquisition parameters and what sort domain your data is in. For application of ravn to make sense this should be shot or group domain.
2. generate the median agc gain surface that ravn is trying to fit a robust surface to. Use the same window length and amplitude parameters and ask for a median agc using a median window increment of 10 milliseconds.
3. make a ravn pass on your data using the default parameters outputting the gain surface only. Examine the relationship between this gain surface and that generated in step 2 above. This surface should look like a robust fit to the former without any of the high frequency information.
4. if the above surface fit seems reasonable then accessing the coefficients file output by the previous run apply the gain surface to your data and examine the results. If you can see no obvious problems then you are ready to apply ravn to your whole dataset.
-
1. determine the optimum -sstep parameter. Reduce this parameter until the best fit to the control agc surface can be found. You may find that past a certain minimum step size the surface changes very little. Use the largest value where this is found to occur. Be aware that the run time for this routine is a direct function of this parameter.
2. scrutinize the coefficients file and determine which order of fit corresponds to the most desirable results. Using the -sord parameter test this order against several others until you are satisfied with the order required to handle your data.
3. if instabilities persist or the fit is unsatisfactory test the -top parameter. Use the same procedure of observing the gain surface until you have determined the best value to use then apply it to your test records.
4. settle on the best -w parameter to use. A decrease in this parameter will decrease the program run time and heat up the scaled output. The optimum window length is decidedly data dependant.
5. once the best result is found from the above procedure, test the -ilim parameter for stability. If you notice very little change in the surface fit with an increase in this parameter then you are in good shape. If large or unexpected changes in the fit occur then your parameterization up to this point is suspect and should be reviewed. A good solution will converge nicely with increasing -ilim. If you cannot stabilize the surface fit by this point in the testing the input data must be ill-conditioned or you really do have large record to record variations in the decay surface [unlikely]. If you cannot do a better job of preconditioning the data then perhaps scaling this data with ravn is not appropriate and another type of scaling should be considered.
6. insert the -median flag on the command line and test the -iter parameter. This option is applicable if your input data displays trace to trace instabilities such as amplifier differences often seen on marine acquisition. Again a good solution will converge with increasing -iter value. Should unpredictable changes occur it is time to revisit a previous parameterization step or question the wisdom of ravn application on your dataset.
Ongoing Experimentation
-
1. Effects of scaling pre or post normal moveout correction on AVO.
2. Case history of AVO effects done on real data with good geologic and engineering control.
Real Dataset Example
What follows is a case history of the preconditioning and parameterization performed on a marine dataset from Nigeria. The displays have been extracted from a quality control subset of records providing single fold coverage over the zone of interest. The dataset was generated using:- editt -N MaipSHOT -R0 25 -O QCdata
Basic Data
The shot records contain 192 traces with a trace spacing of 25 meters. The offset from the source to the near group is 75 meters. The shot interval between records is 25 meters. Examination of the raw data (Fig. 1) reveals significant noise in the water column prior to the first arrivals which must be muted prior to ravn. Traces 67, 93, 100, 131, 179, 180 and 187 are bad and need to be killed throughout. The above was accomplished (Fig. 2) using:
- killtr -N QCdata -K TraceToKill -D -T |\
Interpolate Dead Traces
Since the traces killed, with one exception, leave single trace gaps in the input records it is a trivial task to interpolate data into these traces. The only potential snag is that several of the traces are in the farther offsets where steeply dipping energy is present. To interpolate data into these traces without aliasing these events it is first necessary to apply a pseudo normal moveout correction and linear velocity reduction. Since traces 179 and 180 are adjacent, a different option for interpolation is required. The interpolation was accomplished (fig. 3) using the following flow:
- bdnmo -N QCdataMute -v PseudoVelocity -E |
The reason for the interpolation is to leave no holes in the decay surface sampled by ravn. If large gaps had been present within the shot records the interpolation process may not have been practical and the applicability of ravn would be in question. Once the surface fit coefficients have been derived for the line it is possible to apply them to the non-interpolated data should you so choose.
Attenuate Direct Arrivals
Attenuation of direct arrival energy (Fig. 4) was accomplished using:
- vred -N QCdataInterp -vred 2300 |
NMO Application
In order to provide a less problematic surface to fit, normal moveout was applied using a pseudo velocity function. The function was chosen to reduce the triangular shape of the data within the mute zone while simultaneously minimizing the stretch applied at the far offsets. It is unknown at this writing what affect this approach has on the AVO signature of the data. It may be necessary to use the actual moveout function derived for the line. It may also prove essential to perform the scaling prior to correction for normal moveout. Future experimentation will address these issues with the results being reported in an up-coming USP publication. For now the assumption is made that anything that can be done to minimize the compaction of information on the far offsets will improve the decay surface fit achieved by ravn.To minimize the effects of normal moveout stretch a low cut filter and a post nmo mute was applied. In addition a tn gain correction was performed to reduce the kink in the decay surface along the onset of data. This was accomplished using the following flow:
- bdnmo -N QCdataDirect -v PseudoVelocity |
Quality Control Gain Surface
Prior to ravn parameter testing a median agc gain surface dataset (Fig. 6) was generated. The agc parameters used are those to be used in ravn. The gain surface produced is the surface that ravn will be attempting to calculate a robust fit to. Comparison of the gain surface generated by ravn with this surface serves as a measure for determining optimum parameter values. The following flow was used to generate this surface:
- davc -N QCdataNmo -O QCgainAGC -M -agc -d 10
Parameter Testing
For each parameter test undertaken, both the gain surface and scaled data are produced. The former is contrasted with the above agc surface to gauge goodness of fit. The latter is inspected for undesired scaling effects on the data.Frequently, depending on the color map and scaling used for viewing in xsd, a surface with seemingly good fit produces undesired amplitude variations when applied to the data. This pitfall can be avoided if one uses fixed scaling in xsd to display the data. To obtain the initial multiplier plot the median agc gain surface above using histogram scaling and adjust the percentage until you achieve an acceptable display. Record the scalar and offset applied to the display and use those values to display all gain surface results. Similarly use the scalar and offset determined from the first data result for all future data comparisons.
Defaults
To execute ravn in it's default mode use:
- ravn -N QCdataNmo -O RavnDefault -C CoefsDefault
The resulting gain surface (Fig. 7a) is a demonstrably poor fit to the median agc surface (Fig. 6). A planar surface has been fit to the first 3 records and a fifth order surface fit to the last (Table 1). When applied to the data (Fig. 7b)The deep data is poorly scaled on the first three records while the inside offsets are heavily attenuated on the fourth.
| record | order |
|---|---|
| 126 | 1 |
| 151 | 1 |
| 176 | 1 |
| 201 | 5 |
-tstep
Holding all other command line parameters constant the -tstep entry was varied over the range 1, 4, 16 with the value of 1 yielding the best results. Since this is the default no additional figures are required. A -tstep value of 1 is used for the remainder of the testing.
-sstep
Values of 100, 24 and 12 milliseconds were used. The latter two entries resulted in an improved fit with little difference between them. A value of 12 milliseconds (Fig's 8a, 8b) was chosen for use in the remainder of the testing. Notice that the order of surface determined by the program though increased dramatically (Table 2) is still unstable from record to record. The scaled data, though looking much better at depth, is quite variable in the zone affected by the onset mute.
| record | order |
|---|---|
| 126 | 7 |
| 151 | 5 |
| 176 | 5 |
| 201 | 5 |
-sord
Surface orders of 2, 3, 4, 6, 10 and 13 were tested. A third order surface (Fig's 9a, 9b) was found to give very stable results with a good robust fit to the control gain surface. There is a huge runtime advantage to specifying the surface order [assuming you can for your dataset] as ravn does not have to make this determination by examining several orders. Another way to speed up the process if you cannot hard wire a surface order is to look through your coefficients file to determine the highest order of fit determined during the parameterization tests. Enter this value using the -ord command line parameter to limit ravn to searches of this order or less.
-top
Values of 20, 30, 40 and 50 percent were tested. A value of 30 percent (Fig's 10a, 10b) producing the best result throughout the zone affected by the onset mute was used for the remainder of the testing.
-w
AGC window sizes of 500, 250, 100 and 64 were tested. As the window size was decreased the mean gain surface amplitude was increased. The character of the surface was not changed appreciably. A value of 500 milliseconds, the default, was chosen.
-ilim
This parameters sets the number of iterations of iterative re-weighted least squares to be used in fitting the surface. Values of 3, 5, 7, 11 and 15 were tested. In this case the fit was insensitive to changes in this parameter indicating that the solution had converged by 3 iterations. An ilim value of 3, the default, was chosen.
-median, -iter
Values of 3, 5, 7 and 11 were tested. Again the result was insensitive to changes in the iter parameter. An iter value of 3 (Fig's 11a, 11b) was found adequate to correct for apparent amplifier differences observed on certain channels of the input data. These differences manifest themselves as a slight vertical striping which appears at the same channel locations on all data and so must be a function of the recording system rather than the subsurface.
Conclusions
Through systematic parameter testing a set of processing parameters was derived that allowed the stable application of the ravn process to the entire shot dataset. Failure to undertake this testing would have resulted in a dataset where the interpreter had no idea when the surface fit was adequate and when it was disastrous. As evidenced by the default analysis (Fig's 7a, 7b) an extremely poor fit to the actual decay surface of this dataset would have been used.Preconditioning of the dataset prior to parameterization and application of ravn, coupled with recent changes to the algorithm, particularly the specification of the surface order on the command line, has allowed the establishment of a methodology to address previously noted stability problems.
Return to the USP Home Page.Frame Doc to html conversion by Tony Garossino