Best software for seismic refraction tomography

[Preston09]

I am using SeisImager, but it crashes very often. Any recommendations? Thanks!

[kaminae]

I use zondst2d software. We can compare results of two softwares if You have files with picked times.

[lealoekio]

I heard SeisOpt® @2D™ is very good:
http://www.optimsoftware.com/products/seisopt-2d/

[kaminae]

To make the right choice - we have to check all the software on a same profile. The results can be published in this topic.
I have some good profiles for comparison.

[Preston09]

I like SeisOpt -> requires no prior assumptions of subsurface structure, or any other subjective data, as input.

[kaminae]

Is it good? I think SeisOpt used 1D inversion result at first iteration(You don't see it). Starting from halfspace is impossible for refraction inversion.
At some case ivery important to use a prior model.

[Preston09]

I don't know yet. I only read the introduction and it seems to be more robust than SeisImager.

[kaminae]

Do You have this software to test it with my specific data?

[Preston09]

unfortunately, I don't have it.

[InspecteurColombin]

Hello all,
I am new on this forum and it happens that i have a licence of seisopt@2d. I would be very interested in comparing various inversion algorithms (which also means i do agree to carry out an inversion with seisopt). From my own experience (seisopt@2d and another academic code based on the SIRT technique), the best final image has not always been provided by seisopt. But it remains a very good code which i really like to use.
I confirm that no a priori information is needed to carry on the inversion process. It is however possible to set velocity ranges from travel-time curves analysis. Nevertheless, varying grid sizes have to be tested for each dataset and the code is eventually time-consuming. Batch processing is very much appreciated for this, since it is possible to run the code during the night. The processing time also depends on the chosen resolution (grid cell size) as well as on the number of geophones & shots, but this last point is not a very original information .

[lealoekio]

There is a sample data set at: http://terra.rice.edu/department/facult ... ageep2011/
Apparently it's a big file, but it would be interesting to the inversion results at least using portions/sections of the data set.

[InspecteurColombin]

It appears to be a large very large dataset with up to 10000 picks, which might not be relevant of an "everyday" refraction survey. As you said, we can either use portions of this dataset (1 geophone over 2 and 1 shot over 3, for example) or generate or own synthetic dataset (24 geophones, 2 to 3 layers with varying depths and shapes, etc.).
Since i am on a sick leave with plenty of free time, I can do that and/or sort the sageep dataset. Synthetics have the advantage to be noise and/or human error free and will provide a good estimate of the efficiency of the code, whereas real data will offer a comparison of the robustness of the algorithm to reconstruct the ground structure with picks which may be not perfect.
Otherwise we might use Kaminae's data since he seems to have good profiles.

[Preston09]

Sounds like a plan. I will use the sageep data, probably a section in the beginning of the data to test SeisImage2D as soon as I got time.

[kaminae]

There are inversion result's with zondst2d.
No apriori information. 5 iterations. Time of calculation 7 min.

[kaminae]

Small synthetic data set

[InspecteurColombin]

There are inversion result's with zondst2d.
No apriori information. 5 iterations. Time of calculation 7 min.

I guess it's the result of the inversion of the SAGEEP data set. It is really impressive! Which parameters didi you use (grid cell size for instance) and what is the final RMS error?
On my own, and since I do not have the seisopt@pro licence, I cannot process data with more than 28 shots, i gave up with the sageep data set.
At least, this is an information about which software is the best to invert huge data set between Zond & Seisopt.

I produced a synthetic model as well but I will invert the data you just provided.

[kaminae]

Yes, it's SAGEEP data. As You can see - it is some problem with determination of thin inclined layer. (Which deal with to resolution of refraction tomography).
RMS ~ 1.3% (You can see it in the status bar of program windows). I can use irregular Z and X grid because special forward modeling algorithm is used.
~ 20000 cells for FM and ~3000 for inversion.

Put here Your data set to invert it.

[kaminae]

Layered inversion result

[InspecteurColombin]

OK.
Here is the synthetic model I generated. The details (Vp and Vs) are provided in the image. The 2nd layer stands for a water table (varying Vp but same Vs). There are 9 sources along a 115 m-long profile with 24 geophones spread each 5 m.

3layers.txt

(2.54 KiB) Downloaded 785 times

Following are the results obtained with Seisopt@2d with no a priori information. The grid geometry is 64*44 cells with vertical and horizontal dimensions of 1.8*0.9 m, respectively. The calculation took 25 mn (but I am running the code under windows virtual PC mode which may slow the process down) and I tested 10 configurations (varying number of vertical cells) for a total calculation time of about 3 hours.
The final error, in the RMS sense, is 2%.

One can see that there seem to be a poor lateral continuity around the step. This might come from 1) the seismic substratum is reached and no ray pass below and/or 2) the code produces images with gradients, which is not the case of the model.
Finally, the following provides a hitcount image. It stands for the number of times a cell has been used for the calculation. It mimics a ray-tracing and allows to see where the model is reliable or not.

Since I am not a specialist of inversion, I think it is possible to enhance the results. These images are to be seen as the results of an inversion conducted by the user of a commercial code.

[kaminae]

Results for Your data set.
1. Hodographs.
2. Velocity model.
3. DOI map. Depth of investigation (robustness of result) min is good, max is bad.
4. Model quality. Parameter deal with rays coverage.
5.Layered invesion results.

time of inversion 20 sec!
Little difference between our result as I think deal with different forward solver. As You can see I use smooth velocity model (I have very smooth rays).

[kaminae]

This video demonstrate - how I got result
http://narod.ru/disk/63629785001.1b1916 ... s.zip.html

[InspecteurColombin]

Your data inverted with a SIRT algorithm.
grid: 48*13 cells of 2*2m (1*1 m for the forward computation).
a priori information: homogeneous velocity of 3000 m/s through the whole grid.
9 iterations, 20 s to compute and a final RMS error of 6%.
Interfaces are smoother than yours but it ain't that bad.
With rays:

Same data with a grid cell size of 1*1 m (and 0.5*0.5 m for the forward computation). 15 iterations, less than 60s for the calculation and a final RMS error of 1.8% :

Results of the same data processed with seisopt tomorrow (batch processing during the night...).

[kaminae]

I found some errors in Your synthetic data.
I have modeled right part of section analiticaly for last source. It's simple 1D solution.
And I found big divergence and one strange point. It's similar to addition boundary.
I think may be three cause of error
1. You calculated data for smooth model(and we don't know - how smoothness assigned to model)
2. Model at the picture not corresponded model for calculation.
3. Software error in forward modeling(mesh, or special conditions)

[InspecteurColombin]

You're totally right. I did not check the data... My bad
The error comes from a too large delta x for the forward computation.
Here are the corrected data and I think they are fine.

seismic3layers_fault.txt

(2.66 KiB) Downloaded 830 times

[kaminae]

Ok! It's correct data.
There are results for new file.