Earthquake warning systems are hard, but not having one is worse.

The premise of earthquake early warning systems is simple. An earthquake produces several different kinds of seismic waves that race away from the rupture point. Because they are different kinds of vibrations, they travel at different speeds; and the farther they travel, the more the speedy compressional P-waves pull away from the transverse S-waves, and the more the surface waves lag even further behind.

In the race between different seismic waves, fleet-footed P-waves are heralds for their slower and more earth-shaking brethren.

Fortunately for us, the speediest waves are also the weaker, less damaging ones. The P-waves shake us up a little when they arrive, but they are also giving us a heads up that more damaging shaking is on its way. This warning is at most a few tens of seconds, but with the right infrastructure in place this is enough to shut down vital machinery (trains, elevators, nuclear power stations…) and prepare people for incoming shaking. If detected soon enough, the warning can also be sent ahead of the P-waves at the speed of light, giving even more advanced warning ahead of the expanding front of seismic energy.

Of course, it is much more challenging to put this simple theory into practice. The small window of opportunity for a timely warning can quickly close if the system is not responsive enough. On the other hand, the degree of automation required to gain that responsiveness can lead to a system that is more easily fooled by complex seismic events. Two recent news stories about two of the countries that actually have working earthquake early warning systems highlight challenges from either end of this balancing act.

Mexico: sneak attack from below

Mexico’s earthquake warning system was put into place after a 1985 magnitude 8 rupture on the subduction thrust off the west coast killed thousands in Mexico City. That is the system’s focus: it was built to detect large ruptures on the subduction zone, and warn the residents of Mexico City, who live on top of a massive seismic wave amplifier.

The system worked as designed for the biggest earthquake of 2017, a M8.2 plate bending event. But it struggled to respond quickly enough in the much closer M7.1 a few days later – this NPR story starts with an account of the sirens going off only after the strong shaking started. This earthquake, whilst weaker than the one 12 days earlier, was much closer to Mexico City, resulting in strong shaking that collapsed buildings and killed several hundred people. And that proximity was a problem for the early warning system: with only around one hundred kilometres to travel, rather than several hundred, the P-waves could only pull a little bit ahead of the S-waves and surface waves, leaving barely any time for a warning to get out. The NPR story linked above indicates that changes are already being made to make the warning system more responsive to these kinds of events.

Japan: false positives attack

Last Friday, Japan’s warning system was triggered when it detected P-wave arrivals from what it estimated as a magnitude 6.4 earthquake off the coast north-eastern Japan. No such event had occurred: instead the Japanese Meteorological Agency, who operate the system, reported that the false warning was the result of the early warning system misreading two smaller earthquakes, a M4.4 on the east coast and a M3.9 that occurred on the west coast at the same time, as one larger event.

I was actually interested enough to do a little impromptu data analysis to see if I could work out why the system got fooled. The seismogram for this time from a station in central Japan is a little strange, with very little amplitude variation between the body waves and the surface waves, and earlier P-wave arrivals than expected (a comparison with an M4.7 a little later in the day, in roughly the same location, makes this clear). My speculative interpretation at the time was that the P-waves from the E coast quake reached nearby stations at the same time as surface waves from the smaller, earlier W coast quake. This does seem to have boosted the apparent P-wave magnitude, but by further comparison with the M4.7 seismogram, the boost was clearly not enough to make the signal look like a M6.4. Perhaps it is also a matter of duration: larger ruptures take longer, because a bigger section of the fault is progressively unzipping. If the system interpreted the whole sequence as an extended package of P waves, that may have been sufficient for the system to mistakenly trigger.

Blue: seismogram for the event that triggered Japan’s earthquake early warning system on January 5th, from this station in central Japan. It is probably the hybrid of two events, and looks weird compared to a more normal earthquake (red). Data from IRIS, plotted using Obspy.

Either way, this is a tricky scenario for an automated system to handle, and therein lies the challenge. To save the most lives, people have to respond quickly when an alarm sounds. If you have a computer that cries wolf – a system so sensitive that it is prone to triggering in the absence of a real threat (a false positive) – then people might stop paying attention to it. On the other hand, you don’t want to risk the system failing to trigger when there is a threat (a false negative). This isn’t the first time that Japan’s system has given a false warning – there was also one in August 2016 – but occurrences are hopefully rare enough that the system is still trusted. Even if the alert sound isn’t Godzilla roaring.

Sad update. Asked my friend living in Kyushu who heard plenty of early warning alerts during the Kumamoto EQ and they were not Godzilla sounds.

— Helen Janiszewski (@helenjanisz) January 10, 2018

Canada (and the US): not quite there yet

The problems described above are largely good problems to have, because you actually have a working earthquake warning system in place to struggle with and improve – a system that may not be perfect, but does save lives. On the west coast of North America, despite the looming threat of the San Andreas Fault and especially the Cascadia subduction zone, a functional warning system is still some way from implementation. This article updates the progress on the Canadian side of the border, where ocean bottom sensors and GPS data are being tied into the network to get more timely and accurate detections. I was all ready to use it as a cudgel to whack the US government over the head with for continuing not to properly fund the ShakeAlert system, when I read more closely and realised that the Canadian system is in exactly the same position. They have an at least mostly working prototype, with sensors, and computers dedicated to processing the output of those sensors to generate alerts. But is it the next step, building the infrastructure to get timely warnings out to those in harm’s way, that is the challenging step. Or perhaps more accurately, the challenging step in the US is securing the funding to do so. It’s not cheap ($40 million to set up and $16 million a year to run, the USGS estimates, but it’s a drop in the federal budget to protect 50 million people on the west coast. The lack of urgency is frustrating – perhaps the Canadians will be more sensible.