local-regional case, poor conditioned events, worse than raw data

[saeedsltm]

Hi,

The following is the comparative hypocenter map resulted from running Hypocenter program (left as raw data) and NLLOC (right) after running multiple times with changing VEL2GRID, LOCGRID and LOCSEARCH parameters for defining range of grids from coarse to fine ones. But never could obtained results better than or even similar to raw data. I note that the case suffers from poor conditioned events with small number of P and S phases, large azimuthal gaps and ... but at least I do not expect to get results worse that the raw data. So, my clear question is "Should we expect to get bad results from NLLOC in cases when we encountered in poor conditioned events, even worse than linearized location method like Hypocenter?"

VGGRID 2 501 51 0.0 0.0 -3.0 1.0 1.0 1.0 SLOW_LEN
LOCGRID 401 401 41 -200.0 -200.0 0.0 1.0 1.0 1.0 PROB_DENSITY SAVE
LOCSEARCH OCT 20 20 10 0.001 50000 1000 0 0
LOCMETH EDT_OT_WT 9999.0 4 -1 -1 1.74 -1 -1.0 1

finer grids also were tested, but the results not changed very much.

[alomax]

Hello SaeedSLTM,

Should we expect to get bad results from NLLOC in cases when we encountered in poor conditioned events, even worse than linearized location method like Hypocenter?

In general, the full location information obtained by NLLoc (e.g. the full pdf solution) should be equivalent to and more informative than that of a linearized inversion such as Hypocenter (See Lomax et al. 2001). However, especially in the case of poorly constrained events (e.g. those in the above plots far outsize the station network), the NLLoc maximum likelihood hypocenter may be unstable as it only represents one "optimal" point in the pdf. For locations that are not well constrained, with extensive or irregular location pdf, the expectation hypocenters (STATISTICS ExpectX,Y,Z) may give a more stable map, as it represents (a little bit) better the full pdf.

But here are a few issues I wonder about with your locations:

1. Is Hypocenter fixing the depth or otherwise modifying the used observations or search for poorly constrained events?
2. Is VGGRID 2 501 51 0.0 0.0 -3.0 1.0 1.0 1.0 SLOW_LEN a good representation of the position of layer boundaries in your model, or are there layers that are finer than 1 km or which do not fall on integer km depths? Maybe try an equivalent VGGRID based on a cell size of 0.1 km.
3. What values are you using for P and S pick uncertainties and LOCGAU, LOCGAU2 travel-time uncertainties? If these values are unreasonably large, then the NLLoc pdf's are enlarged, smoothed and less precise, with possible increased scatter in maximum likelihood hypocenters. Conversely, if the uncertainties are too small the pdf can become complex and splintered. Also, EDT_OT_WT may effectively reject picks with large residuals as outliers, if the pick uncertainties are set unrealistically small. The uncertainties should also be roughly equivalent to the effective pick and travel-time uncertainties used by Hypocenter, in order to compare the two methods.

I hope the above helps some, there are likely other reasons for the differences in epicenters...

Best regards,
Anthony

Lomax, A., Zollo, A., Capuano, P., & Virieux, J. (2001). Precise, absolute earthquake location under Somma-Vesuvius volcano using a new three-dimensional velocity model. Geophysical Journal International, 146(2), 313–331. https://doi.org/10.1046/j.0956-540x.2001.01444.x

[saeedsltm]

Dear @alomax

Thanks for your answer.

Is Hypocenter fixing the depth or otherwise modifying the used observations or search for poorly constrained events?

Actually no, the depths are not fixed during relocation, but there might be some depths where some events have the similar values, and this is a very common issue with linearized location programs like Hypocenter. Especially in poorly constrained cases.

Is VGGRID 2 501 51 0.0 0.0 -3.0 1.0 1.0 1.0 SLOW_LEN a good representation of the position of layer boundaries in your model, or are there layers that are finer than 1 km or which do not fall on integer km depths? Maybe try an equivalent VGGRID based on a cell size of 0.1 km.

I've noticed as I followed up your comments on Issues pages. Since the velocity model that we used is a coarse one (5 layers) with all integer thicknesses, so I think we don't need to create finer grid than 1.0 km grid point spacing. BTW, I event tested 0.1 km grid spacing and no significant changes resulted.

LAYER 0.00 5.40 0.00 3.10 0.00 2.70 0.00
LAYER 4.00 5.70 0.00 3.28 0.00 2.70 0.00
LAYER 8.00 6.15 0.00 3.53 0.00 2.70 0.00
LAYER 11.00 6.40 0.00 3.68 0.00 2.70 0.00
LAYER 19.00 6.90 0.00 3.97 0.00 2.70 0.00
LAYER 46.00 8.05 0.00 4.63 0.00 2.70 0.00

What values are you using for P and S pick uncertainties and LOCGAU, LOCGAU2 travel-time uncertainties?

According to relocation results obtained by Hypocenter the average residual for all stations lies within a range of -0.6s - 0.6s and the average RMS for 5600 events is ~ 0.4s. The following is the considered parameters:
LOCGAU 0.2 0.0
LLOCGAU2 0.02 0.2 2.0

So, I think these values are reasonable according to mid distances between events and stations and the calculated residuals.

Best