Toward a Three-Dimensional Crustal Structure of the Dead Sea region from Local Earthquake Tomography (2004)

Ph.D. thesis, Tel-Aviv University, pp. 123.

F. Aldersons

Abstract

We studied the local seismicity of the Dead Sea basin for the period 1984-1997. Sixty percent of well-constrained microearthquakes (ML ≤ 3.2) nucleated at depths of 20-32 km and more than 40 percent occurred below the depth of peak seismicity situated at 20 km. With the Moho at 32 km, the upper mantle appeared to be aseismic during the 14-year data period. A relocation procedure involving the simultaneous use of three regional velocity models reveals that the distribution of focal depths in the Dead Sea basin is stable. The lower-crustal seismicity under the basin is not an artifact created by strong lateral velocity variations or data-related problems. An upper bound depth uncertainty of ± 5 km is estimated below 20 km, but for most earthquakes depth mislocations should not exceed ± 2 km. A lithospheric strength profile has been calculated. Based on a surface heat flow of 40 mWm-2 and a quartz-depleted lower crust, a narrow brittle to ductile transition might occur in the crust around 380°C at a depth of 31 km. For the upper mantle, the brittle to ductile transition occurs in the model at 490ºC and at 44 km depth. The absence of micro-seismicity in the upper mantle remains difficult to explain.

Within the framework of local earthquake tomography, a new automatic picking system called MannekenPix has been developed in order to collect a high-quality set of picking data from the Dead Sea region. In another work, this data set would then become the input of a high-resolution travel-time tomographic study of the crustal structure of the basin. In the first step of MannekenPix, the seismograms are filtered by a high-fidelity Wiener filter in order to increase the signal-to-noise ratio of the P-wavetrain before picking. The Wiener filter of MannekenPix uses the maximum entropy method to estimate power spectral densities, a method particularly well adapted to short data segments. The picking is performed in the second step by the robust and versatile Baer-Kradolfer (Baer and Kradolfer, 1987) picking engine. If a valid pick has been found, a variable delay correction is then applied in the third step in order to reduce the inherent delay of the Baer-Kradolfer algorithm arrivals. In the fourth step, the weighting engine provides a statistical estimate of the uncertainty. For each data set, the weighting engine has to be calibrated first by a multiple discriminant analysis performed on user-supplied reference picks and weights.

By using MannekenPix to pick the seismograms from well-constrained earthquakes of the Dead Sea region, 526 (99.1 %) out of 531 earthquakes are locatable by automatic picking. Out of 15,250 seismograms, 6,889 (45.2 %) were routinely picked and 7,089 (46.5 %) were automatically picked. On the calibration data set, the average discrepancy between the routine picking and the reference picking is 0.037 sec, with a standard deviation of 0.121 sec. The average discrepancy of the automatic picking is +0.007 sec with a standard deviation of only 0.066 sec. For the complete data set, 3,058 phases (43.1 %) of the 7,089 automatic P picks fall into predicted weight 1 (absolute uncertainty not greater than 40 msec), 1,173 (16.5 %) fall into class 2 (absolute uncertainty between 40 msec and 80 msec), 1,554 (21.9 %) fall into class 3 (absolute uncertainty between 80 msec and 140 msec) and 1,304 (18.4 %) fall into class 4 (absolute uncertainty greater than 140 msec). The results from a totally out-of-sample set of reference picks and weights confirm the stability of predicted weights on unseen data.

Di Stefano et al. (2004, in revision) present the application of MannekenPix to about 240,000 seismograms from nearly 29,000 local earthquakes routinely recorded by the Italian national seismic network during the period 1998-2001. MannekenPix was able to produce 103,131 P picks (73% of the 139,500 routine onsets) from 23,108 events out of 28,900 events. About 17,130 phases (17 %) of the 103,131 fall into class 1 (absolute uncertainty not greater than 0.1 sec) and 15,429 (15 %) fall into class 2 (absolute uncertainty between 0.1 sec and 0.2 sec). Results show that MannekenPix arrival times for classes 1 and 2 combined, produce a distribution of residual times with a much smaller standard deviation than for the CSI bulletin. This means that a significant increase in the quality of the data set has been gained, coming at the price of a reduced quantity. The application of MannekenPix to the very large and inhomogeneous dataset of waveforms recorded in Italy from 1998 to 2001 yields nearly 33,000 high-quality weighted picks and polarities. By using MannekenPix we achieved our goal to build a new dataset of P-wave arrival times and polarities belonging to user-defined classes 1 and 2 (the highest qualities) with associated error estimations. The analysis of relocation residuals and the time versus distance plots for these classes show the effectiveness of the picking system. Although the hit rate of MannekenPix 1.6.2 applied to this large and noisy dataset did not exceed 75 % of the rate of an expert seismologist, the consistency of computed arrivals and polarities and their rapid estimation are very suitable to substitute bulletin data, solving the typical problem of extending consistency and quality to large datasets, and saving a great amount of time. We believe that the improved seismic dataset is suitable to refine the seismic tomography and to improve the knowledge of the local and regional stress fields in the Italian peninsula through the determination of a large number of focal mechanisms for well-located events, extending the analysis to small magnitudes and past events.

 


Full Text (PDF 4.0MB)