The Canterbury, NZ, Sequence

Transcription

1 The Canterbury, NZ, Sequence Complex: M7.1 Darfield (Sep 10) M6.2 Christchurch (Feb 11) M6.0 Christchurch (Jun 11) M5.9 Christchurch (Dec 11) Devastating: Over 180 deaths $10-15 billion USD Raised expected hazard: Time-dependence Gerstenberger et al. (2014), Earthquake Spectra Wealth of data to study earthquake clustering and predictability. 1

2 Retrospective Evaluation of Earthquake Forecasts during the Canterbury, New Zealand, Earthquake Sequence Max Werner, M. Gerstenberger, M. Liukis, W. Marzocchi, D. Rhoades, M. Taroni, J. Zechar, C. Ca>ania, A. Christophersen, S. Hainzl, A. Helmste>er, A. Jimenez, S. Steacy & T. Jordan University of Bristol, GNS Science, SCEC/USC, INGV, GFZ Potsdam, ISTerre Grenoble, University of Ulster, University of Adelaide StatSei9 2015, Potsdam 2

3 Collaboratory for the Study of Earthquake Predictability Cyber-infrastructure for blind, prospective and retrospective assessment of forecasting models in a variety of tectonic environments Global 13 models Western Pacific 16 models EU Testing Center SCEC Testing Center China Testing Center ERI Testing Center Zurich Beijing California 86 models Los Angeles Italy 48 models Oceanic Transform Faults 1 model North-South Seismic Belt Tokyo Japan 229 models GNS Science Testing Center Wellington Testing Center Testing Region CSEP Testing Regions & Testing Centers Upcoming Upcoming 451 models under test in April, 2015 New Zealand 58 models

4 Retrospective Canterbury Experiment RetrospecRve evaluaron of the predicrve skills of earthquake forecasrng models during the sequence Goals: Improve our understanding of earthquake clustering Evaluate newly developed physics- based and hybrid starsrcal/physical against starsrcal models Characterize influence of real- Rme data on forecast quality Help guide model development for OperaRonal Earthquake ForecasRng (OEF) 4

5 Experiment Design ForecasRng target: Number of earthquakes per 0.05 by 0.05 degree cell in tesrng region, per 0.1 magnitude unit bin, above magnitude M 3.95 per forecast horizon Several forecast horizons and evaluaron periods Start right a"er M7.1 Darfield earthquake Forecast horizons: 1- yr, 1- mo, 1- day EvaluaRon periods: Non- overlapping, but Re- generate forecasts right a"er large quakes (Feb 11, June 11, Dec 11) Forecasts with 1- yr horizon also evaluated on enrre horizon Two modes Retrospec/ve, using best data available today as input Pseudo- prospec/ve, using preliminary data as input

6 15 Installed Models Coulomb-based Statistical clustering Smoothing Type Features Names Authors Physical Coulomb & Dieterich s (1994) rate- state fricron CRS0 (original + stress change uncertainres), CRS1 (+ receiver geometry uncertainres), CRS2 (+ available ajershock FMs) CRS3 (+ all ajershocks as sources) CRS4 (+ sparally- variable stressing rate) Ca>ania et al. Sta/s/cal ETAS ETAS 0 (epicenter) & ETAS 2 (fault geometry) Hainzl et al. Space- Rme smoothing K2 (kernels, with Gutenberg- Richter) K3 (kernels, with non- parametric magnitude) Helmste>er & Werner SpaRal Coulomb & STEP STEP- Coulomb Steacy et al. SpaRal Coulomb & ETAS ETAS 1 Hainzl et al. ETAS & producrvity RETAS 0 (epicenter) & RETAS 2 (fault) Hainzl et al. ETAS & producrvity & SpaRal Coulomb RETAS 1 Hybrids Reference Uniform Poisson SUP et al. Uniform PoissonHainzl model Rhoades

7 Data Sets Target data set to evaluate forecasts: Best available earthquake catalog (394 eqks. M 3.95) Geonet 4 September March 2012 Input/training data set for model input 1. Mode 1 (retrospecrve) Best available earthquake catalog from Geonet Best available slip models (Beaven et al., 2012) Best available focal mechanisms 2. Mode 2 (pseudo- prospecrve) Near- real- Rme earthquake catalog snapshots (captured by NZ CSEP TesRng Center) Slip models within first 10 days (Holden et al., 2011 & personal comm.) 7

8 1-yr forecasts after M7.1 Darfield Coulomb/rate-state model (CRS-0) Cattania et al. Hybrid STEP-Coulomb model Steacy et al. Testing period: Target events: 2010/09/ /09/02 M 3.95 (328) incl. M6.2 Feb and M6 June Christchurch eqks. 8

9 1-yr forecasts after M7.1 Darfield ETAS-2 CRS-3 Testing period: Target events: 2010/09/ /09/02 M 3.95 (328) incl. M6.2 Feb and M6 June Christchurch eqks.

10 I N (A,B) = 1 N I N : N i=1 " $ # ln A i B i % ' N A N B & N rate- corrected informaron gain N: number of observed earthquakes A, B: forecasts to compare A i N A : rate of forecast A in bin i total rate of forecast A Information Gains: T & W Tests Is I N significantly different from 0? apply Student s paired t-test: T-test apply Wilcoxon signed rank test: W-test

11 Results: T/W tests 1-yr forecasts Statistical model Hybrid model Physical model best-available input data Prob. Gain 60/eqk real-time input data Testing period: 2010/09/ /02/29 2 x 1-yr forecasts (second forecast scaled) Target events: M 3.95 (394) Reference model: Uniform Poisson 11

12 Results: T/W tests 1-yr forecasts Prob. Gain 180/eqk Statistical model Hybrid model Physical model Testing period: 2010/09/ /02/29 4 x 1-yr forecasts reset after major eqks. Target events: M 3.95 (394) Reference model: Uniform Poisson 12

13 Results: T/W tests 1-mo forecasts Prob. Gain 735/eqk Statistical model Hybrid model Physical model Testing period: 2010/09/ /02/29 20 x 1-mo forecasts reset after major eqks. Target events: M 3.95 (394) Reference model: Uniform Poisson 13

14 Results: T/W tests 1-day forecasts Prob. Gain 3000/eqk Statistical model Hybrid model Physical model Testing period: 2010/09/ /02/29 1-day forecasts reset after major eqks. Target events: M 3.95 (394) Reference model: Uniform Poisson 14

15 Discussion In 1992 M7.1 Landers sequence CSEP experiment, Coulomb models could not compete (Woessner et al., 2011) what changed?? Random luck? Improved (over- ) fiong? Improved model components/choices? Does Coulomb/rate- state mechanism apply more to Canterbury? Real- Rme data degrades forecasts, right? Not necessarily: strongly model- dependent Newly developed models: Some improved Coulomb/rate- state models With uncertainres in receiver geometry and ajershocks as stress sources. Hybrids perform mixed STEP- Coulomb best model for 1- yr (w/ resets), 1- mo, 1- day Non- parametric models lack magnitude- dependence Probability gains of up to 3,000 per eqk. 15

16 Conclusions This CSEP experiment quanrfied the predicrve skills of Rme- dependent models, which can reach probability gains of up to 3,000 per eqk. over Rme- independent seismicity hazard models during acrve sequences. In contrast to the Landers CSEP experiment, we find that the predicrve skill of physics- based models can exceed that of starsrcal models. Time- stamped, versioned earthquake parameter catalogs are required to further study influence of real- Rme data on forecast quality. 16