Embracing Data Incompleteness for Better Earthquake Forecasting

The magnitude of completeness (Mc) of an earthquake catalogue, above which all earthquakes are assumed to be detected, is of crucial importance for any statistical analysis of seismicity. Mc has been found to vary with space and time, depending in the longer term on the configuration of the seismic network, but also depending on temporarily increased seismicity rates in the form of short-term aftershock incompleteness. Most seismicity studies assume a constant Mc for the entire catalogue, enforcing a compromise between deliberately misestimating Mc and excluding large amounts of valuable data.

Epidemic-Type Aftershock Sequence (ETAS) models have been shown to be among the most successful earthquake forecasting models, both for short- and long-term hazard assessment. To be able to leverage historical data with high Mc as well as modern data, which is complete at low magnitudes, Leila Mizrahi, Shyam Nandan and Stefan Wiemer developed a method to calibrate the ETAS model when time-varying completeness magnitude Mc(t) is given. This extended calibration technique is particularly beneficial for long-term Probabilistic Seismic Hazard Assessment (PSHA), which is often based on a mixture of instrumental and historical catalogues.

In addition, the researchers designed a self-consistent algorithm which jointly estimates ETAS parameters and high-frequency detection incompleteness, to address the potential biases in parameter calibration due to short-term aftershock incompleteness. For this, they generalized the concept of Mc and consider a rate- and magnitude-dependent detection probability – embracing incompleteness instead of avoiding it.

To explore how the newly gained information from the second method affects earthquakes' predictability, Mizrahi, Nandan and Wiemer conducted pseudo-prospective forecasting experiments for California. Two features of our model are distinguished: small earthquakes are allowed and assumed to trigger aftershocks, and ETAS parameters are estimated differently. The researchers compare the forecasting performance of a model with both features and two additional models, each having one of the features to the current state-of-the-art base model.

Preliminary results suggest that their proposed model significantly outperforms the base ETAS model. They also find that the ability to include small earthquakes for the simulation of future scenarios is the main driver of the improvement. This positive effect vanishes as the difference in magnitude between newly included events and forecast target events becomes large. Mizrahi, Nandan and Wiemer think that a possible explanation for this is provided by the findings of previous studies, which indicate that earthquakes tend to preferentially trigger similarly sized aftershocks. Thereby, besides being able to better forecast relatively small events of magnitude 3.1 or above, the researcher gained a useful insight that can guide them in developing of the next, even better, earthquake forecasting models.

The paper "The Effect of Declustering on the Size Distribution of Mainshocks" is published here:

This work has received funding from the European Union’s Horizon 2020 research and innovation program under Grant Agreement Number 821115, real‐time earthquake risk reduction for a resilient Europe (RISE).


An Energy‐Dependent Earthquake Moment–Frequency Distribution

The probability of a strong shock to nucleate inside a small area A that just experienced another strong event, should be lower than the probability predicted by the tapered Gutenberg-Richter (TGR) model. This is because a lot of elastic energy has already been released in the first strong shock, and it takes time to recover to the previous state. However, this fact is not taken into account in the TGR model. Due to this limitation, RISE members Ilaria Spassiani and Warner Marzocchi propose the energy-varying TGR (TGRE) model. Here, the researchers impose the corner seismic moment to be a space-time function depending on the amount of elastic energy E currently available in a small area A. More precisely, their energy-varying corner seismic moment Mc (E) increases with the square of the time elapsed since the last resetting event, which is supposed to have reset the elastic energy in the small area A to a minimum value.

In practice, the taper of TGRE is abruptly shifted to the left just after the occurrence of a strong shock, and then it slowly recovers to the long-term value with the energy-reloading process. Spassiani and Marzocchi impose TGRE to verify an invariance condition: when considering large domains, where a single strong event cannot significantly affect the whole energy available, TGRE becomes the TGR. A sensitivity analysis also shows that the dependence of Mc (E) on its parameters is not substantial in the short-term, proving that their specific choice, as long as reasonable, cannot affect the results of any analysis involving TGRE.

To evaluate the reliability and applicability of the TGRE model, the two RISE members applied it to the Landers sequence (USA) which started with the Mw7.3 mainshock that occurred on the 28th of June 1992. First, they found out that the seismic moment-frequency distribution (MFD) close to the fault system affected by the mainshock is statistically-significantly different from that of earthquakes off-faults, showing a lower corner magnitude. This result is entirely independent of modeling, and it underlines the need for an energy-varying MFD at short spatiotemporal scales. Then, Spassiani and Marzocchi have shown that TGRE may explain well the difference in the MFDs for the Landers sequence, and that the results are stable for possible variations of its parameters. Notably, they obtained positive evidence in favour of the TGRE fit, with respect to TGR, for Landers data within one week, one month, six months and one year since the Mw7.3 mainshock (Fig. 1 shows the results for one week). Most of the times, the evidence is “substantial” and “strong” (terminology by Kass & Raftery, 1995).

The results suggest that TGRE can be profitably used in operational earthquake forecasting, as the model is simple and rooted in clearly stated assumptions. It requires some more or less explicit subjective choices. However, Spassiani and Marzocchi think they are less subjective than ignoring the empirical evidence that shows that strong triggered earthquakes do not nucleate on a fault just ruptured by another strong event. Rather than focusing on the details of the model, which is obviously not the only one possible, the RISE members mainly aim to get across the message of including an energy dependency in MFD: Spassiani and Marzocchi claim that self-organized criticality at large spatiotemporal scales changes in intermittent criticality at small space-time domains recently experiencing a significant release of energy.
This study has been recently published on BSSA:

As future work, Spassiani and Marzocchi aim at evaluating the TGRE reliability and the comparison with alternative models through prospective tests.

This study was partially supported by the Real‐time Earthquake Risk Reduction for a Resilient Europe (RISE) project and funded by the European Union’s Horizon 2020 research and innovation program (Grant Agreement Number 821115).


Dynamic seismic risk communication and rapid situation assessment in face of a destructive earthquake in Europe

On December 29th 2020, a strong earthquake with a magnitude of 6.4 shook the city of Petrinja, 45 km south of the capital city of Zagreb. Eight people lost their lives. In March 2020, eight months before the Petrinja event, an earthquake with a magnitude 5.3 damaged the capital Zagreb to a similar extend. Due to the Zagreb earthquake and its aftershocks, the LastQuake app developed at the Euro-Mediterranean Seismological Centre (EMSC) had a high penetration rate of 7 % in the country at the time of the Petrinja earthquake. In addition, EMSC had released a new website for mobile devices to optimise crowdsourcing of felt reports a few weeks before the Petrinja earthquake occurred. These specific circumstances made the Petrinja earthquake an ideal test case for dynamic risk communication and the rapid situation assessment tools developed by EMSC within the RISE and TURNkey projects.

Both projects aim at reducing future human and economic losses caused by earthquakes in Europe. TURNkey (Towards more Earthquake-resilient Urban Societies through a Multi-sensor-based Information System enabling Earthquake Forecasting, Early Warning and Rapid Response actions) is a project funded under the same H2020 call as RISE.

The Petrinja mainshock was detected within 66 seconds through a surge of EMSC website traffic and 11 seconds later from concomitant LastQuake app openings. These crowdsourced detections initiated a so-called “Crowdseeded seismic Location” (CsLoc). It is a method that exploits the geographical information of crowdsourced detection to select the seismic stations likely to have recorded the earthquake. The method also uses the time information to identify observed arrival times likely associated with this earthquake to perform fast and reliable seismic location. The automatic CsLoc was published 98 seconds after the earthquake and it was located 8 km away from the location that was manually reviewed afterward. EMSC received more than 2'000 felt reports in the first 10 minutes, out of a total of more than 15'000, before its servers started to experience significant delays due to the high traffic. These initial felt reports were sufficient to evaluate the epicentral intensity at VIII (for a final epicentral intensity of IX), i.e. damaging levels. Furthermore, they confirmed the heads-up from the automatic impact assessment tools "EQIA" (Earthquake Qualitative Impact Assessment). About 140 geo-located pictures were crowdsourced showing the effects of the earthquake. EMSC shared all the information on rapid situation assessment with the ARISTOTLE (All Risk Integrated System Towards Trans-boundary holistic Early-warning) group in charge of preparing an impact assessment for the European Civil Protection Unit within 3 hours of the event.

A posteriori determined rupture orientation and position from felt reports collected within the first 25 minutes using the FinDer software developed at ETH Zurich are in good agreement with fault orientation from tectonic settings and aftershock distribution. This approach will be tested further in operational conditions in the coming months.

The large volume of collected felt reports accelerated the cooperation with the United States Geological Survey (USGS) to integrate them in USGS global ShakeMaps and define common geographical clustering approaches of individual reports.

Beyond these scientific results, a significant effort was devoted to complete automatic information published on Twitter @LastQuake by answering questions and explaining the reasons for service disruptions observed after several main aftershocks occurred. All communication was performed in English. We retweeted many of the tweets from the Croatian seismological institute and referred to them as much as possible. Questions followed according to the time evolution already observed after other damaging earthquakes. First, people were interested in the expected impact, then in the possible evolution of the seismicity, in the earthquake prediction (with the traditional confusion between forecast and early warning), and finally whether human activities (in this case hydrocarbon exploitation) could have caused the tremblors. During this period which lasted about two weeks, there were also many questions about seismology, such as the meaning of the magnitude, intensity, or the reasons for magnitude discrepancies between institutes.

It is difficult to evaluate the impact of this direct public communication effort quantitatively. Due to the English, in fact, only part of the population was reached. However, there are elements that seem to indicate that Twitter users appreciated our effort. The number of tweets viewed reached 9 million the day of the mainshock, and an average of 4.8 million over the first seven days. Again, this demonstrated the strong interest of the public in receiving information after a damaging earthquake and during an aftershock sequence. Tweets explaining that earthquake prediction does not exist were liked 700 times. Many users reported that getting rapid information and direct answers to their questions was the key to decrease their anxiety. This public interest led to tens of interviews in national and regional media. But perhaps more meaningful in terms of public appreciation, EMSC collected more than 2000€ from individual donations and received many offers for technical support in relation to our service disruptions from both IT specialists and companies. 

This experience also underlined the many improvements that still need to be done, from strengthening our IT infrastructure to hierarchize information during an aftershock sequence better. In general, the great interest shown by the public in Croatia in receiving information after the Petrinja event demonstrates a strong need for dynamic risk communication. Moreover, it proves that crowdsourcing can significantly improve the capacity for rapid impact assessment.


Authors: Rémy Bossu, Istvan Bondàr, Maren Böse, Jean-Marc Cheny, Marina Corradini, Laure Fallou, Sylvain Julien-Laferrière,  Frédéric Massin, Matthieu Landès, Julien Roch, Frédéric Roussel and Robert Steed


What is the Foreshock Traffic Light System?

Recently, Laura Gulia investigated the spatio-temporal evolution of the earthquake size distribution throughout a seismic sequence focusing on the b-value. This is a parameter characterizing the relationship between the earthquake magnitude and the number of earthquakes. She and her colleagues found out that, immediately after a mainshock, the b-value increases by 20%-30% and remains high for at least the following 5 years, reducing the chance of occurrence of a larger earthquake near the fault that originated the mainshock.

Based on their research, Laura Gulia and Stefan Wiemer developed the Foreshock Traffic Light System (FTLS), a promising tool for the mainshock and aftershock hazard assessment. In the interview with Gabriele Amato from the NH blog, Laura Gulia explains more about the thoughts behind it and future research plans. Read the full interview here.

Gulia, L., and S. Wiemer (2019). Real-time discrimination of earthquake foreshocks and aftershocks, Nature 574, 193–199.


Looking into the future of forecasts

In the United Kingdom’s House of Commons in 1854, a Member of Parliament stood up and made the suggestion that recent scientific advances might allow the weather in the city to be known ‘twenty-four hours in advance’. The House broke into uproar and laughter - the idea was considered utterly preposterous. But with thousands of lives being lost in the country every year as a result of storms, by 1861 storm warnings were being wired to ports using the new telegraph system. So popular were they, that these ‘weather forecasts’ quickly became a staple part of newspaper content across the country.

Now, 160 years later, operational earthquake forecasting is in a similar position. With a proliferation of sensors that would have been considered infeasible perhaps 50 years ago alongside growing computer modelling power and expertise, geoscientists increasingly have information about potential seismic activity that could be of use in emergency and public planning. But how best to communicate that?

Over a century of experience in communicating the risk of life-threatening storms has put meteorology in a strong position to help us tackle this problem, but they are not the only ones: those used to communicating flooding, epidemics of disease and even financial market fluctuations all have lessons we can learn from.
As well as talking to communications professionals in all these fields, we are also listening to people ‘on the ground’ in three key RISE countries: Italy, Switzerland and Iceland. By interviewing members of the public, emergency responders and long-term planners and testing our messages and visualisations on them we will hopefully soon be able to advise on how best to get useful ‘earthquake forecasts’ into the hands of those who can act on them.
Hopefully as well as learning from the successes of weather forecasting, we can be prepared by its failures. Tragically the father of the weather forecast, Robert Fitzroy, beset with scepticism from scientific colleagues about his methods, funding problems from government, and complaints from those who lost business as a result of false alarms in the warnings, killed himself before he saw them become the ubiquitous and lifesaving service that they are today. With the backing of RISE and alert to these potential barriers, we hope to overcome them.


A closer look #3: Structural Health Monitoring – Opportunities for Integrating Sensing Data into Rapid Loss Assessment

The extreme loads imposed by earthquakes threaten the integrity of the built environment. As not all buildings react in the same way to earthquakes, a rapid understanding of the extent of damage to buildings and its consequences on providing safe shelter for the population is a crucial contribution to an earthquake-resilient Europe. Therefore, in a similar way to doctors who examine vital functions to diagnose the health of their patients, structural health monitoring allows engineers to diagnose the integrity of buildings.

In the absence of means for direct measurements of building damage, one objective of the RISE project consists in finding indirect indicators of damage. Data-driven structural health monitoring uses damage-sensitive indicators, which are derived from the building’s earthquake response providing a real-time performance indication. To this end, signal processing, statistical analysis and machine learning are used to derive performance indicators from the time-and-frequency domain representation of the response. The increasing availability of sensing hardware at low cost, combined with the ever-growing possibilities for local data processing offered by the Internet-of-Things capabilities, provide exciting opportunities towards smart structures, which support engineers and decision-makers in the immediate aftermath of earthquakes. Hence, the early response to earthquake events can be improved by comparison to the current practice of time-consuming and potentially subjective visual inspections.
Well-designed damage-sensitive indicators help to more precisely diagnose damage by providing higher-level information regarding the location and the severity of building damages. The RISE project, through the breadth of its network, offers a rare opportunity to combine building-specific values from structural health monitoring with regionally applicable building behaviour models. With the engineering knowledge of building taxonomies and damage accumulation, the automation provided by data-driven structural-health monitoring can enable rapid assessment of regional consequences to the built environment, induced by earthquake events, and further provide guidance for rapid recovery. 

Related sites


International Day for Disaster Risk Reduction #DRRday

The United Nations General Assembly has designated October 13th as the International Day for Disaster Risk Reduction to promote a global culture of disaster risk reduction. Disasters have a vital impact on people's lives as well as on their wellbeing. Concerning catastrophes caused by nature such as thunderstorms, landslides, or droughts, earthquakes are the deadliest natural hazard. However, many damages and losses can be avoided through effective risk reduction strategies.

In March of this year, an earthquake with a magnitude of 5.4 occurred near Zagreb (Croatia). Bricks fell from roofs, facades cracked, walls collapsed, debris damaged parked cars, and many citizens got injured. Up to now, earthquakes cannot be predicted, but there are measures and approaches for minimizing their consequences. Developing tools and measures to reduce future human and economic losses is the aim of RISE, which stands for "Real-time earthquake rIsk reduction for a reSilient Europe". RISE studies seismic risk, its changes, importance, and evolution at all stages of the risk management process. The project depicts current potentials and limits as well as advances the state of the art to reduce seismic risk in Europe and beyond.

After one year of the project, first results are becoming apparent. RISE contributes in many ways to gain knowledge, which will be beneficial to reduce further earthquake-related losses. For instance, RISE has successfully deployed a prototype array as a demonstrator in Bern (Switzerland) and designed an impulse generator, which is currently being tested in multi-storey buildings. In addition, the researcher focused on developing new and extending existing approaches to model seismicity. Some models have already demonstrated, and therefore, the project has made notable progress in the fields of earthquake forecasting. It also focused on physics-based modelling of seismicity, an evolving field. Furthermore, static and time-invariant exposure models for 45 countries and time-invariant vulnerability models representing over 500 buildings have been developed. To ensure the best possible usage of all available information for the benefit of society, RISE scientists tested different start page designs and hazard announcements representing the diversity of elements used in multi-hazard platforms and conducted workshops to understand which features of multi-hazard warning apps non-experts prefer.

Thus, RISE adopts an integrative, holistic view of risk reduction targeting the different stages of risk management. Improved technological capabilities are applied to combine and link all relevant information to enhance scientific understanding, inform societies and consequently foster Europe's resilience and beyond.

More information:



4 new publications in "The Power of Citizen Seismology: Science and Social Impacts"

Several RISE papers (some of them co-funded by the Turnkey project) dealing with crowdsourced data and public earthquake communication have recently been published. Four papers are part of the special issue of the open-access online Frontiers journal “The Power of Citizen Seismology: Science and Social Impacts” edited by Rémy Bossu, Kate Huihsuan Chen and Wen-Tzong Liang.

These four RISE papers published in the Frontiers journal cover interactions with the public, how seismological community can benefit from such interaction for rapid situation awareness and a method to speed-up seismic locations of felt earthquakes. You can find all papers here: 

Are you interested in more RISE publications? Here you will find the list with all our publications.


*The image for this news item is adapted from: Bossu et al. (2020), "Rapid Public Information and Situational Awareness After the November 26, 2019, Albania Earthquake: Lessons Learned From the LastQuake System"



A closer look #2: What is Operational Earthquake Forecasting?

One important focus of RISE is to advance earthquake predictability research such as Operational Earthquake Forecasting. This research can benefit from the constantly evolving observational capabilities of seismic monitoring efforts, which, for instance, result in an ever-increasing amount of recorded earthquakes, especially toward smaller magnitudes. Such capabilities need to be exploited to gain more insight into the earthquake occurrence processes and, therefore, to improve earthquake forecasting.

In our first step, we explore existing high-resolution earthquake catalogues that contain events with magnitudes down to ML0 or below. We started to develop an interactive tool that will facilitate and aid us in a more intuitive analysis of seismicity in five dimensions (see Figure).

In particular, we will focus on these aspects:

  1. Spatio-temporal variability in the frequency-magnitude distribution: e.g., statistical analyses of event sizes could tell us more about the state of a fault system.
  2. Earthquake clustering properties: e.g., well-located hypocenters could reveal how earthquake sequences progress and how earthquakes are triggered.
  3. Foreshock analysis: e.g., earthquakes prior to a larger earthquake might share a common spatial-temporal pattern. In addition, high-resolution catalogues could potentially reveal many more sequences that have foreshocks than is currently believed.
  4. Limits of the current quality of earthquakes catalogs, e.g., what information are we missing?

We will adopt state-of-the-art methods (e.g., from the machine learning domain) to augment these analyses, for instance, to employ a parameter selection and search for signals and patterns that are indicative of the earthquake occurrence process.

Our findings will have an impact on improving our understanding of the earthquake occurrence process. Our gained knowledge could allow us to develop innovative earthquake forecasting models, which can be stochastic, physics-based and/or of a hybrid type. Ultimately, our advances will contribute to mitigating better the seismic risk, which will be analysed within another work package of RISE.


A closer look #1: Towards optical sensing of ground motion for improved seismic hazard assessment

Optical fibres are the backbone of our modern communication network. Short pulses of laser light transmit enormous amounts of data, but on their journey from sender to the receiver, they also gather information about the optical fibre itself. In fact, microscopic displacements of the fibre slightly distort the laser pulses – an effect that has recently become detectable with highly sensitive interferometers.

This emerging technology, known as Distributed Acoustic Sensing (DAS), allows us to measure ground motion excited by a large variety of sources, such as earthquakes or landslides. Harnessing existing networks of telecommunication fibres, DAS, therefore, offers the opportunity to assess and potentially mitigate natural hazards in densely populated urban areas.

To explore this opportunity, RISE researchers at ETH Zurich are conducting a pilot experiment in the Swiss capital Bern, closely collaborating with the telecommunication company SWITCH. Several connected telecommunication fibres are traversing the city in different directions along with a 6 km long path measure ground motion every two metres, in real-time, nearly 1000 times per second. Most of the observed ground motion is caused by traffic, industrial installations, and construction sites.

Though the amplitude of these signals is, fortunately, much lower than the ground motion caused by destructive earthquakes, this wealth of data can be used to infer rock properties of the upper tens to hundreds of metres of the subsurface. Knowing these properties is essential to predict the ground motion caused by potential future earthquakes.

Research on DAS in urban environments is in its infancy, within the RISE project and worldwide. Initial results are very promising, especially in terms of the quality and unprecedented spatial resolution of the data. Yet, substantial research and development are still needed to process the enormous amounts of DAS data efficiently.


Detect earthquake-triggered landslides via Twitter

Detect earthquake-triggered landslides via Twitter

Landslides may significantly hamper earthquake response, because they can block roads. However, there is very little information about how, where and when earthquakes triggered landslides within the first few hours after a global earthquake occurred. Therefore, the European-Mediterranean Seismological Centre (EMSC) started a project in collaboration with the Qatar Computing Research Institute (QCRI) to evaluate whether it is possible to detect such landslides by monitoring tweets on Twitter in real time. 

Although such detection already exists for earthquakes at EMSC, the detection of landslides presents a different challenge, as the number of tweets published on this topic is very small. In order to tackle this issue, EMSC is working together with QCRI, which has developed AIDR (Artificial Intelligence Disaster Response). This is a platform, which harvests tweets from Twitter and uses artificial intelligence (AI) to analyse the pictures. A database of landslide pictures was the first training dataset generously provided by the British Geological Survey. Further AI training is currently ongoing. Besides that, we are manually checking all detected tweets on a daily basis and hope to have an operational prototype in the coming months.


Promising Kick-Off Meeting

Promising Kick-Off Meeting

The RISE kick-off meeting took place from 2 to 4 September 2019 in Zurich, Switzerland. Over 60 participants joined to discuss the project's first steps. Besides many interesting presentations held project members and external experts, a poster fair and breakout sessions provided enough room to be creative and to take everyone's expertise and ideas into account. During the conference dinner on a boat on Lake Zurich, the participants had the possibility to network and discuss in a different setting. Now the RISE community is ready to start with their work to make Europe more resilient.


RISE kick-off meeting in Zurich

RISE kick-off meeting in Zurich

From 2 to 4 September 2019, the RISE kick-off meeting will be held in Zurich. Over 60 representatives of the scientific community are expected to attend. The event will feature several poster sessions, breakout sessions, and a networking event to support the projects collaboration.