Another problem with using the Nevada Geodetic Laboratory products is that they apply a symmetrical smoothing filter to the data, so the large signals of the coseismic displacements are partially smoothed backwards in time. Angelyn Moore pointed this out.
It is also not clear whether the observation in the lab and that derived from GPS are comparable. Lab-based slip measurement often relies on a single site near one edge of the fault. It is implicitly assumed that the measured slip is homogeneous over the entire fault, which has been shown invalid by array- or digital-image-correlation(DIC)-based slip measurement. On the other hand, GPS measures the deformation at the Earth's surface, which could sample a large volume of deformation signals at depth (cross-talk effect, convolution effect, etc.).
Nice article. Reminds me the paper below. They suggest there was a months-long thousand-kilometre-scale precursor before the Tohoku. I vaguely remember folks on Twitter were as well questioning about whether that was noise? Or maybe, those common mode noises are indeed precursory signals, but we just don't understand what it is?
Bedford, J.R., Moreno, M., Deng, Z., Oncken, O., Schurr, B., John, T., Báez, J.C. and Bevis, M., 2020. Months-long thousand-kilometre-scale wobbling before great subduction earthquakes. Nature, 580(7805), pp.628-635.
Hi Baoning, thanks for the pointer to that paper! I recall reading it when it came out. The wobbles they are looking at took place over periods of several months. The data were all processed to remove network-level common mode noise, so the signal they are interpreting would likely be more similar to our common-mode corrected time series. But Bedford et al. went much farther in their analysis and had more advanced methods for looking at longer timescale signals. It does seem reasonable to suggest that GPS wobbles indicative of slow slip over length scales of a thousand kilometers should be cause for concern, even if there is not yet enough data to say whether these large wobbles are common precursors of really big earthquakes. But it is also interesting to think about the practical issues involved if a 3-month periodicity precursor signal is every confidently detected. How would you go on alert effectively for half a year - or maybe longer???
I'm not seeing where you demonstrate that the 12-hour period in their fig 2d is not real. Does it also badly fade when properly denoised? It is physically plausible - faster slip when the tides are encouraging, although this assumes the earthquakes themselves initiate with a strong correlation with the tidal stress, which is probably not right.
Hi John, the denoised data does fade the periodic signal - there are still ups and downs, but not clearly with a standard period. (There is a plot showing the denoised global stack in the post.) We did not calculate a periodicity on this time series. Since there was a M7.3 earthquake just before the Tohoku timeseries (which contributes a big chunk of the stack), with continued seismicity during the time considered, these up-and-down variations are not surprising.
If you check out our follow-up post (https://earthquakeinsights.substack.com/p/update-on-apparent-gps-detection) you'll see a plot that shows the global stack using only the common-mode noise. That time series visibly contains the same 12-hour period, indicating that whatever the source of the 12-hour periodic signal, it does not appear to be the later earthquakes, but something more fundamentally related to the GPS data overall (for instance, tides).
Aha, now I notice that your "Global stack following per-earthquake common mode noise removal" plot is the one I should have studied. It does seem to have lost the near-12-hour cycle. Hard to be sure, as the amplitude is only slightly diminished (you should use the same vertical axis scaling to make comparison easier) but in any case what peaks might be present are no longer with the same phasing.
I'm not seeing anything left in the stacks to interpret, except to quantify the lack of precursory slip.
Great analysis! Except you sucked me into a rabbit hole of geodesy and data analysis when I should be digitizing a geologic map or posting YouTube videos. I do notice the Nevada Geodetic Laboratory both offers a 1.5-hour latency GPS product and claim to apply a "Stochastic Kalman filter/smoother" (http://geodesy.unr.edu/gps/ngl.acn.txt); if they're using that latency to smooth their data in near-real time, that could be where the just under 2 hour upswing comes from . . . but I'm no geodesist . . . so I'll just desist.
Yes, Eric Fielding mentioned the symmetrical smoothing filter. I suppose it could be partially responsible for the signal, much of which comes from Tohoku. The 2-hour upswing is still present in the far-field data, but even those data might be affected by Tohoku since "far field" is >200 km from the hypocenter, not from the eventual rupture - many of the stations would be close enough to the actual rupture to be affected. I imagine it would be pretty easy to find out by looking at the original data, but I'm also not a geodesist!
What difference would it make if all the gps stations were moving, more or less the same? I note Japan is a small piece of a much larger tectonic plate being under-slid by the Pacific plate.
I note also Japan is just one example, what about others, do they all suffer the same flaw?
That is basically what is happening in several of the networks - it's making up most of the apparent signal. It looks like that "signal" is mostly coming from a few networks (although there are 90 earthquakes, 8 of them contribute 2/3 of the time series, since only a few places have dense networks). We are still exploring some of the details. For many of the networks, it is not possible to calculate the error, because there are not enough stations to do a reliable calculation. For instance, 22 of the earthquakes were only recorded by a single high-rate GPS station.
Another problem with using the Nevada Geodetic Laboratory products is that they apply a symmetrical smoothing filter to the data, so the large signals of the coseismic displacements are partially smoothed backwards in time. Angelyn Moore pointed this out.
I didn't know that - thanks! Do you know what the time window of the smoothing is?
It is also not clear whether the observation in the lab and that derived from GPS are comparable. Lab-based slip measurement often relies on a single site near one edge of the fault. It is implicitly assumed that the measured slip is homogeneous over the entire fault, which has been shown invalid by array- or digital-image-correlation(DIC)-based slip measurement. On the other hand, GPS measures the deformation at the Earth's surface, which could sample a large volume of deformation signals at depth (cross-talk effect, convolution effect, etc.).
Nice article. Reminds me the paper below. They suggest there was a months-long thousand-kilometre-scale precursor before the Tohoku. I vaguely remember folks on Twitter were as well questioning about whether that was noise? Or maybe, those common mode noises are indeed precursory signals, but we just don't understand what it is?
Bedford, J.R., Moreno, M., Deng, Z., Oncken, O., Schurr, B., John, T., Báez, J.C. and Bevis, M., 2020. Months-long thousand-kilometre-scale wobbling before great subduction earthquakes. Nature, 580(7805), pp.628-635.
Hi Baoning, thanks for the pointer to that paper! I recall reading it when it came out. The wobbles they are looking at took place over periods of several months. The data were all processed to remove network-level common mode noise, so the signal they are interpreting would likely be more similar to our common-mode corrected time series. But Bedford et al. went much farther in their analysis and had more advanced methods for looking at longer timescale signals. It does seem reasonable to suggest that GPS wobbles indicative of slow slip over length scales of a thousand kilometers should be cause for concern, even if there is not yet enough data to say whether these large wobbles are common precursors of really big earthquakes. But it is also interesting to think about the practical issues involved if a 3-month periodicity precursor signal is every confidently detected. How would you go on alert effectively for half a year - or maybe longer???
Oh I see. Bedford et al.'s signals do seem to be more robust then. Thanks, Kyle!
Nice work!
Yes, when removing noise removes the interpreted signal, the conclusion is clear.
I'm not seeing where you demonstrate that the 12-hour period in their fig 2d is not real. Does it also badly fade when properly denoised? It is physically plausible - faster slip when the tides are encouraging, although this assumes the earthquakes themselves initiate with a strong correlation with the tidal stress, which is probably not right.
Hi John, the denoised data does fade the periodic signal - there are still ups and downs, but not clearly with a standard period. (There is a plot showing the denoised global stack in the post.) We did not calculate a periodicity on this time series. Since there was a M7.3 earthquake just before the Tohoku timeseries (which contributes a big chunk of the stack), with continued seismicity during the time considered, these up-and-down variations are not surprising.
If you check out our follow-up post (https://earthquakeinsights.substack.com/p/update-on-apparent-gps-detection) you'll see a plot that shows the global stack using only the common-mode noise. That time series visibly contains the same 12-hour period, indicating that whatever the source of the 12-hour periodic signal, it does not appear to be the later earthquakes, but something more fundamentally related to the GPS data overall (for instance, tides).
Aha, now I notice that your "Global stack following per-earthquake common mode noise removal" plot is the one I should have studied. It does seem to have lost the near-12-hour cycle. Hard to be sure, as the amplitude is only slightly diminished (you should use the same vertical axis scaling to make comparison easier) but in any case what peaks might be present are no longer with the same phasing.
I'm not seeing anything left in the stacks to interpret, except to quantify the lack of precursory slip.
Good point about the vertical axis scaling! It's definitely a bit hard to keep track of what's what in these time series.
Great analysis! Except you sucked me into a rabbit hole of geodesy and data analysis when I should be digitizing a geologic map or posting YouTube videos. I do notice the Nevada Geodetic Laboratory both offers a 1.5-hour latency GPS product and claim to apply a "Stochastic Kalman filter/smoother" (http://geodesy.unr.edu/gps/ngl.acn.txt); if they're using that latency to smooth their data in near-real time, that could be where the just under 2 hour upswing comes from . . . but I'm no geodesist . . . so I'll just desist.
Yes, Eric Fielding mentioned the symmetrical smoothing filter. I suppose it could be partially responsible for the signal, much of which comes from Tohoku. The 2-hour upswing is still present in the far-field data, but even those data might be affected by Tohoku since "far field" is >200 km from the hypocenter, not from the eventual rupture - many of the stations would be close enough to the actual rupture to be affected. I imagine it would be pretty easy to find out by looking at the original data, but I'm also not a geodesist!
What difference would it make if all the gps stations were moving, more or less the same? I note Japan is a small piece of a much larger tectonic plate being under-slid by the Pacific plate.
I note also Japan is just one example, what about others, do they all suffer the same flaw?
That is basically what is happening in several of the networks - it's making up most of the apparent signal. It looks like that "signal" is mostly coming from a few networks (although there are 90 earthquakes, 8 of them contribute 2/3 of the time series, since only a few places have dense networks). We are still exploring some of the details. For many of the networks, it is not possible to calculate the error, because there are not enough stations to do a reliable calculation. For instance, 22 of the earthquakes were only recorded by a single high-rate GPS station.
Fascinating!