Complex world of quake forecasting
The 7.8 magnitude Kaiko¯ura Earthquake that struck just after midnight last November 14 was a monster that tore great scars in the landscape and ripped up the seabed.
Thankfully, it wasn’t followed by a terrifying procession of major aftershocks.
There was the harrowing magnitude 5.7 quake that hit particularly hard the town of Scargill in North Canterbury on November 22, setting residents in the northeast of the South Island on edge again.
That turned out to be one of only two quakes classed as ‘severe’ by GeoNet in the past year that didn’t happen on November 14. The other was a 5.5 magnitude quake centred near Seddon on December 4.
A month after the Kaiko¯ura quake, GeoNet published an article, which said the aftershock sequence was ‘‘falling within or just below the lower end of our forecast range’’.
It helped that the bottom of the ranges for different size quakes and time periods was often zero.
In its latest aftershock forecast - for the two months from October 19 - GeoNet puts the number of 5.0-5.9 magnitude quakes expected at from 0-6.
For magnitude 6.0-6-9 and magnitude 7.0 and above, the range is 0-1.
GeoNet makes it very clear these are forecasts.
They are probabilities of earthquakes happening in a specified magnitude range, spatial area and certain time period.
They are not predictions giving specific times and locations. ‘‘At present there is no scientific way to accurately and
‘‘At present there is no scientific way to accurately and reliably predict when and where a big earthquake is going to happen next.’’
reliably predict when and where a big earthquake is going to happen next,’’ GeoNet said.
It is using observations of previous earthquake patterns in New Zealand and around the world and the vast amounts of data from the November 14 earthquake to try to improve the forecasting of aftershock sequences following future big quakes.
GNS Science hazard and risk scientist Dr Annemarie Christophersen said the Kaiko¯ura aftershock sequence ‘‘seems a bit less productive than the average’’ New Zealand aftershock sequence.
If some people felt the aftershock sequence following Kaiko¯ura was less active than after the Canterbury earthquakes, one reason might be that many Canterbury aftershocks happened in urban areas where many people felt them, Christophersen said.
Also, the Canterbury sequence was invigorated with further large aftershocks about every six months for a couple of years.
To complicate the counting of earthquakes above a certain size for comparison of the Kaiko¯ura and Canterbury sequences, the magnitude of earthquakes of some sizes used to be slightly overestimated before the introduction of an automated earthquake processing system in 2012.
Obviously aftershock forecasting is complicated. The models used to forecast aftershocks use previous earthquakes as input, and are updated as new earthquakes happen.
For the best results, all earthquakes would have ideally been processed in the same way, and all earthquakes above a certain threshold size would be recorded. But that’s not the case. After a big earthquake, the seismic network gets overloaded and smaller earthquakes are initially not detected.
In a paper written for the New Zealand Society for Earthquake Engineering on operational earthquake forecasting (OEF), Christophersen and colleagues at GNS Science said careful filtering and seismic processing allowed some of the smaller quakes to be detected.
‘‘However, this is a timeconsuming manual process. For example, it took about 18 months for the first 24 hours to be processed following the M7.1 Darfield earthquake.’’
Secondly, earthquake magnitude has been measured in different ways, and there have been step changes in the way earthquakes have been processed in New Zealand.
For example, from the mid 1980s until 2011 a seismic processing system called CUSP was used, then in 2012 GeoNet introduced the more automatic processing system called SC3.
‘‘We are still doing work to understand the effect of the change of the processing software on the magnitudes,’’ Christophersen said.
Another wrinkle is that the recording of smaller earthquakes has improved over the decades. The density of seismic stations started to increase in the 1960s, and from that point earthquakes of about magnitude 4.0 and above could generally be detected.
If that doesn’t make things difficult enough, the magnitude of the November 14 earthquake was upgraded from M7.5 to M7.8 two days after the event.
The OEF paper said that for one simple aftershock model that implied the expected number of earthquakes doubled in any given time period.
Under the initial magnitude, the number of aftershocks had agreed ‘‘reasonably well’’ with the number of aftershocks forecast, but the number of actual earthquakes fell below the forecast when the magnitude went up.
‘‘It is not clear whether the initial agreement between the data and the model is due to incomplete detection of many large aftershocks, or whether the sequence is much less productive than the average New Zealand sequence,’’ the paper said.
Christophersen said even taking into account any M5-M5.9 quakes missed in the first couple of days, the aftershock sequence still seemed a bit less productive than average.
‘‘But we do not understand the difference in magnitude well yet that might explain a large component of that difference,’’ she said.
The GNS scientists are developing hybrid models for forecasting earthquakes at short, medium and long time periods.
The latter is also known as time-invariant and is used for such things as engineering design codes.
Basically the models rely on two types of earthquake clustering in time and space: the decay of aftershocks following a large earthquake, and the increase of seismicity building up before a large earthquake.
Models are developed by fitting data from past earthquakes. The models are then tested against other data to see how well they work.
‘‘We have learned that hybrid models that combine different sources of data ... work better than models that are based on one concept and source of data only,’’ Christophersen said.
‘‘We still try new data sources and ways of combining them, and then test them against each other and the data.’’
Research done in New Zealand has also found that the rate and size of smaller earthquakes can increase over years or decades before a large earthquake. That’s the clustering before a large earthquake mentioned above.
A model, known as EEPAS, based on that observation has been developed by Dr David Rhoades of GNS Science. It is being tested in this country and other earthquake-prone areas around the world. Two versions of EEPAS are part of the Kaiko¯ura hybrid forecast.
‘‘The EEPAS model has been applied to a number of regional earthquake catalogues and consistently forecasts major earthquakes better than timeinvariant models,’’ the OEF paper said.