Hi Kyle and Judith, thanks for this =compelling and accessible forensic exploration of both the data and proposed underlying processes put forward by Bedford et al. (2020).
You mention a lack of comparison with other geodetic data. However, similar ideas have been put forward and discussed in a series of papers focused on anomalies in GRACE gravity in the months prior to the same earthquakes.
I also hope your post will stimulate more discussion and analysis.
Hi Roland, Judith and I really appreciate your comment on this post. It seems like there are dozens of papers on pre-Tohoku precursor-like behaviours, running the full gamut of observations from gravity to GPS to seismicity, etc. We have been thinking about looking at the pre-Tohoku gravity change papers more closely. They fall pretty far outside our previous experience in terms of methodology, but I guess that means we stand to learn a lot if we read the papers all carefully! Cheers.
Hello, a very complete work but I have two doubts.
Firstly in figure 21, you show a signal fitting, which did not find the jump, and then you say: "It seems like the first run of the GrAtSiD algorithm did not discover the jumps on October 20 and November 20, 2010. This is not unexpected: the figure above shows that it can take several iterations of the algorithm to reliably find all of the jumps. The jumps on those dates are also small compared to the data scatter, and easy to miss." You don't considerate that they used a algorith to find jump before the first convergence, this can be a reason why you didn't find the jump.
Bedford et al. (2020) say: "the steps imposed owing to known hardware changes or when the station is within a radial cut-off distance from a catalogued earthquake, with the cut-off distance r (in kilometres), being r = 10^(0.5M − 0.8), where M is the earthquake magnitude (as also used by the Nevada Geodetic
Laboratory).
Second, also Bedford et al. (2020) say: "The purpose of the initial GrAtSiD fit is to estimate a residual time series for each component of each station. Using the common-mode error reduction approach, we then took the median residual value for each directional component (as a function of time) and subtracted this from the corresponding time series. Next, we applied GrAtSiD again on the common-mode corrected time series. The common-mode filter also serves as a low-pass filter and is effective in removing high-frequency noise such as reference frame jitter as well as reducing heteroskedasticity in the network solutions". They filter the residuals with a low-pass filter, you don't say that, only a median of the stack. This process can help with the noise in the HKWS, PIMO, GUUG and KIRI stations (figure 13).
It is impossible to remove all the noise from the data, but we do what we can with the data we have. Thank you! I have learned from your article, good luck.
Regarding your first point: the jumps are clearly visible in Figure 2a of the original paper. The caption for that figure says: "Offsets (steps) in these time series (automatically solved by GrAtSiD) have been removed." So, the algorithm did not find and remove the jumps - they were not removed before or during use of the GrAtSiD algorithm. When we looked at the data available in the supplement, we can see the jumps that were estimated and corrected by GrAtSiD, and there are none on those two dates. Even if some other method was applied to correct additional jumps, it is not relevant as the jumps remain in the "dejumped" time series.
As a note: Figure 21 in our post is not data from Japan, it is an illustration of how GrAtSiD works with synthetic data from their original 2018 paper. The purpose of the figure is not to highlight what happened specifically in Japan, but to illustrate how GrAtSiD works through iterations to identify signals in the time series.
Regarding your second point: the authors do not apply a low-pass filter specifically. Instead, they say that one of the effects of the common-mode filter (which we describe) is that it also serves as a low-pass filter (removing high frequency noise).
I hope that clarifies these points! Thanks for reading the article. Certainly we hope that people continue to investigate geodetic signals to understand tectonics; however, it is important to remember that these datasets contain unexplained noise that is difficult or impossible to remove without additional information. I hope that more work is done to understand the source of this noise to allow more detailed research.
Hi Kyle and Judith, thanks for this =compelling and accessible forensic exploration of both the data and proposed underlying processes put forward by Bedford et al. (2020).
You mention a lack of comparison with other geodetic data. However, similar ideas have been put forward and discussed in a series of papers focused on anomalies in GRACE gravity in the months prior to the same earthquakes.
I also hope your post will stimulate more discussion and analysis.
Panet, I. et al (2018). https://doi.org/10.1038/s41561-018-0099-3
Wang, L., & Bürgmann, R. (2019). https://doi.org/10.1029/2019GL082682
Bouih, M. et al. (2022). https://doi.org/10.1016/j.epsl.2022.117465
Panet, I. et al. (2022). https://doi. org/10.1029/2022JB024542
Hi Roland, Judith and I really appreciate your comment on this post. It seems like there are dozens of papers on pre-Tohoku precursor-like behaviours, running the full gamut of observations from gravity to GPS to seismicity, etc. We have been thinking about looking at the pre-Tohoku gravity change papers more closely. They fall pretty far outside our previous experience in terms of methodology, but I guess that means we stand to learn a lot if we read the papers all carefully! Cheers.
Quite interesting, thanks for taking a close look and framing it wrt the regional tectonics.
Hello, a very complete work but I have two doubts.
Firstly in figure 21, you show a signal fitting, which did not find the jump, and then you say: "It seems like the first run of the GrAtSiD algorithm did not discover the jumps on October 20 and November 20, 2010. This is not unexpected: the figure above shows that it can take several iterations of the algorithm to reliably find all of the jumps. The jumps on those dates are also small compared to the data scatter, and easy to miss." You don't considerate that they used a algorith to find jump before the first convergence, this can be a reason why you didn't find the jump.
Bedford et al. (2020) say: "the steps imposed owing to known hardware changes or when the station is within a radial cut-off distance from a catalogued earthquake, with the cut-off distance r (in kilometres), being r = 10^(0.5M − 0.8), where M is the earthquake magnitude (as also used by the Nevada Geodetic
Laboratory).
Second, also Bedford et al. (2020) say: "The purpose of the initial GrAtSiD fit is to estimate a residual time series for each component of each station. Using the common-mode error reduction approach, we then took the median residual value for each directional component (as a function of time) and subtracted this from the corresponding time series. Next, we applied GrAtSiD again on the common-mode corrected time series. The common-mode filter also serves as a low-pass filter and is effective in removing high-frequency noise such as reference frame jitter as well as reducing heteroskedasticity in the network solutions". They filter the residuals with a low-pass filter, you don't say that, only a median of the stack. This process can help with the noise in the HKWS, PIMO, GUUG and KIRI stations (figure 13).
It is impossible to remove all the noise from the data, but we do what we can with the data we have. Thank you! I have learned from your article, good luck.
Hi Lorenzo, thanks for your comments!
Regarding your first point: the jumps are clearly visible in Figure 2a of the original paper. The caption for that figure says: "Offsets (steps) in these time series (automatically solved by GrAtSiD) have been removed." So, the algorithm did not find and remove the jumps - they were not removed before or during use of the GrAtSiD algorithm. When we looked at the data available in the supplement, we can see the jumps that were estimated and corrected by GrAtSiD, and there are none on those two dates. Even if some other method was applied to correct additional jumps, it is not relevant as the jumps remain in the "dejumped" time series.
As a note: Figure 21 in our post is not data from Japan, it is an illustration of how GrAtSiD works with synthetic data from their original 2018 paper. The purpose of the figure is not to highlight what happened specifically in Japan, but to illustrate how GrAtSiD works through iterations to identify signals in the time series.
Regarding your second point: the authors do not apply a low-pass filter specifically. Instead, they say that one of the effects of the common-mode filter (which we describe) is that it also serves as a low-pass filter (removing high frequency noise).
I hope that clarifies these points! Thanks for reading the article. Certainly we hope that people continue to investigate geodetic signals to understand tectonics; however, it is important to remember that these datasets contain unexplained noise that is difficult or impossible to remove without additional information. I hope that more work is done to understand the source of this noise to allow more detailed research.