2011 SCEC Annual Meeting: Workshop on Transient Anomalous Strain Detection Jessica Murray-Moraleda, U. S. Geological Survey Rowena Lohman, Cornell University September 11, 2011
From the SCEC3 Science Objectives: “Develop a geodetic network processing system that will detect anomalous strain transients” Systematic monitoring lagged despite Growth in permanent GPS and strainmeter networks InSAR time series analysis techniques Growing number of transient events observed world-wide SCEC CMM 4.0
From the SCEC3 Science Objectives: “Develop a geodetic network processing system that will detect anomalous strain transients” Transient detection algorithms enable Real-time monitoring of transient deformation and associated seismicity Characterization of signals for investigating underlying processes Identification of non-tectonic signals Tracking of data quality Planning future network development to improve detection thresholds. SCEC CMM 4.0
What is a transient? – Real-time monitoring of transient deformation and associated seismicity – Characterization of signals for investigating underlying processes – Identification of non-tectonic signals – Tracking of data quality – Planning future network development to improve detection thresholds. Previous efforts: – Require some spatial, temporal coherence – “Characterization” requires treatment of seasonal + new, larger postseismic Retrospective vs. real-time analysis – Discuss today Target audience? – Make progress before AGU Key issues:
Dragert et al. (2001) 15 days 5 mm Cascadia slow slip event recorded in east component at ALBH Slow slip event in Manawatu region of the North Island of New Zealand Wallace and Beavan (2006) 2.5 cm 7 mos. Transients vary in duration and amplitude
Transients often propagate spatially The signal may only be apparent on a small number of sites at any given time. Space-time history of Cascadia slow slip events Szeliga et al., 2008)
Data are contaminated by time-varying non-tectonic signals Bennett (2008) King et al. (2007) Above right: time-varying seasonal signal; Below: spatially-coherent hydrologic signal
Apparent transients can signal site-specific problems Figures courtesy of Tom Herring (MIT) Above: seasonal trend actually due to a malfunctioning antenna; Left: apparent transient due to snow on the antenna.
August 2008 Held a brainstorming workshop. Framed the problem. Identified test exercise as preferred approach to foster an active community of researchers explore promising methodology combine effective approaches in novel ways. Debated use of real versus synthetic data. August 2008 Held a brainstorming workshop. Framed the problem. Identified test exercise as preferred approach to foster an active community of researchers explore promising methodology combine effective approaches in novel ways. Debated use of real versus synthetic data.
September 2008 Announced Transient Detection Exercise at SCEC Annual Meeting. September 2008 Announced Transient Detection Exercise at SCEC Annual Meeting.
January 2009 Began Phase I. Established group websites for file exchange (data, results, true signals) and discussion. January 2009 Began Phase I. Established group websites for file exchange (data, results, true signals) and discussion.
March 2009 Phase I results submitted. High SNR case Primarily used for validating code March 2009 Phase I results submitted. High SNR case Primarily used for validating code
June 2009 Phase II data released. June 2009 Phase II data released.
August 2009 Phase II results submitted. All high SNR signals were detected and well-characterized. Low SNR signals were almost universally undetected. August 2009 Phase II results submitted. All high SNR signals were detected and well-characterized. Low SNR signals were almost universally undetected.
September 2009 Workshop held at the SCEC annual meeting. September 2009 Workshop held at the SCEC annual meeting.
October 2009 Phase IIC data released. October 2009 Phase IIC data released.
December 2009 Convened special session “Detection and Characterization of Transient Crustal Deformation” at Fall AGU meeting. December 2009 Convened special session “Detection and Characterization of Transient Crustal Deformation” at Fall AGU meeting.
February 2010 Phase IIC results submitted. February 2010 Phase IIC results submitted.
August 2010 Phase III results submitted. August 2010 Phase III results submitted.
September 2010 workshop September 2010 workshop
December 2010 AGU special session “Development and Testing of Methods for Detecting and Estimating Unsteady Motion in Geodetic Time Series” in conjunction with Simon Williams. December 2010 AGU special session “Development and Testing of Methods for Detecting and Estimating Unsteady Motion in Geodetic Time Series” in conjunction with Simon Williams.
Phase IV data released and examined by subset of groups
Maria Liukus implements Lohman test algorithm in testing center
Todays workshop
AGU workshop
Results obtained using different algorithms: Successful at retrospectively detecting signals already visible in time series Less successful with subtle signals Real-time capabilities not yet assessed Need to establish detection thresholds as a function of signal magnitude spatial extent duration network configuration Need to quantify the false alarm rate (will be easy once codes are “detection- center-ready) In Phase III, onwards, participants reported confidence on detections, generally in a qualitative sense. Outcomes from previous workshops
Refinements to be made to synthetic test data: Data covariance: The noise spectra of the test data can be assessed using statistical methods without the data covariance, but the covariance provides information about bad data Simulating the data covariance structure for synthetic data will require further examination of error statistics More subtle signals More realistic signals such as offsets (coseismic or instrumental) spatially-coherent non-tectonic transients postseismic (of various mechanisms) Action: Phase IV workshops had a variety of signals, with seasonal variations that varied from year to year. No postseismic transients yet. Non-tectonic transients included in PhaseIII went mostly undetected.
Outcomes from previous workshops Synthetic versus real test data: Synthetic data is useful for trouble-shooting and improving algorithms makes assessing success easier embodies assumptions about signals encourages “tuning” of algorithms to anticipated signals Consensus: There is substantial additional source complexity yet to be added to the synthetic time series, and algorithms are still early in development, so continue to focus on synthetic test data Since last year: Duncan’s code was made freely available Three new datasets appeared to be examined by only three groups Barriers?
Some thoughts on future directions for transient detection Who will use detection algorithms? How will they use them? What is the optimal level of physics that should be brought to bear? Is it enough to identify that a change is taking place? What range of signal characteristics can one algorithm be expected to detect? To what extent should the algorithm be expected to classify the source? How should we quantify the level of certainty at which a detection is made? How do these requirements vary depending on user? For real data - how do we deal with the large, known transients at Parkfield, Mojave and now in the Salton Trough?