Time picking and random noise reduction on microseismic data

Methods based on the STA/LTA ratio and the modified energy ratio (MER) are studied for their efficacy in automatic first-arrival time picking on high-noise microseismograms. Testing of the two methods on both field and synthetic data indicate that they can pick arrival times for signal-to-noise (SNR) levels as low as 1.7, but that MER time picking yields more consistent results. Moreover, MER time picking is significantly faster than STA/LTA time picking. For many seismic processing procedures, reducing the random noise on the input seismograms will lead to much improved results. Trace averaging (i.e., stacking) over a group of seismograms aligned in phase is a standard method for reducing random noise. Time shifts for trace alignment can be found using one of two methods: finding the average trace that gives minimum variance to the shifted input traces, and using the time picks from the MER method. A noise-signal separation (NSS) technique, developed as an extension to simple trace averaging, separates a noisy seismogram into a noisereduced signal component and a random noise component while preserving the relative signal amplitudes on the input seismograms. INTRODUCTION Arrival-time picking is a fundamental and ubiquitous process in the seismic industry. For example, with seismic surface surveys, we need first-arrival picks to derive the velocity information on near-surface low-velocity zones for making static corrections. With full waveform sonic (FWS) logs, we need first-arrival picks to derive the velocity information on formation intervals encountered by the logging tool as it moves up or down the well (Han et al., 2008). In the analysis of microseismic data, arrival-time picking on microseismograms must be done before hypocenter location can be done (Han et al., 2009; Wong et al., 2009b). Many seismic processing and imaging methodologies are limited by random noise in the raw field seismograms. In the microseismic case, the hodogram/back-azimuth method for hypocenter location using low input seismograms with signal-to-noise ratios (SNR) leads to unacceptably large uncertainties in the microseismic source coordinates. To improve the end product of these processing procedures, random noise on field seismograms must be attenuated. We have developed a random noise suppression methodology that involves trace averaging and a subsequent noise-signal separation (NSS) step. TIME PICKING ON MICROSEISMIC DATA Arrival time-picking for direct P and/or S events is a critical step for achieving the primary goal of microseismic monitoring and analysis, namely, estimating hypocenter locations to track fracture growth during reservoir stimulation by hydraulic fracturing. In other applications (Maxwell and Urbancic, 2001), microseismic data from a passive Han, Wong, and Bancroft 2 CREWES Research Report — Volume 21 (2009) seismic monitoring system may be recorded every 15 seconds for a period of days, months, or even years. Therefore, a microseismic dataset may consist of tens or hundreds of thousands of seismic traces, even if the vast majority is discarded after being classified as containing no useful information. FIG. 1: Definition of the STA/LTA method. The length of the LTA energy collection window is five to ten times the length of the STA energy collection window, which needs to be on the order or two to three times the dominant period of the seismic arrival. FIG. 2: Definition of the MER method. Preceding and following energy collection windows with equal lengths are located at a test point. The window lengths are two to three times the dominant periods of the seismic arrival. For handling such large numbers of traces, the time-picking method must be fast, accurate, and automatic. We have developed the modified energy ratio (MER) method as Microseismic noise reduction and time picking CREWES Research Report — Volume 21 (2009) 3 a faster alternative to the STA/LTA method widely used for microseismic and earthquake analysis. Figure 1 shows the definition for the STA/LTA ratio (Chen and Stewart, 2005; Munro, 2004). Figure 2 shows the definition of the MER method. Both methods rely on energy ratio calculations, differing only on the lengths and positions of energy-collecting windows and the details of ratio evaluation at each test point. For the STA/LTA method, the time pick occurs at the maximum of the rising slope of the ratio (or, equivalently, for digitized plots, at the maximum of the numerical derivative). For the MER method, the time picks occurs at the peak of the MER attribute. Comparison of different modified energy ratios Equation 1 defines the basic energy ratio attribute: eeee(ii) = ∑ ggeegg(jj)2 ∕ ∑ ggeegg(jj)2 ii−LL jj=ii ii+LL jj=ii , (1) where i is the test point index, L is the length of the energy collection windows preceding and following the test point, and grm (j) is the seismogram value at index j. The basic energy ratio can be modified by combining with the absolute values on the seismogram in different ways: eeee1(ii) = eeee(ii) · aaaaaa(ggeegg(ii)) , (1a) eeee2(ii) = eeee(ii) · [aaaaaa(ggeegg(ii))]3 , (1b) eeee3(ii) = [eeee(ii) · aaaaaa(ggeegg(ii))]3 . (1c) FIG. 3: Comparison of different modifications of the basic energy ratio for various noise levels. The modifications improve the ability for the attribute to detect the onset of a seismic arrival in the presence of random noise. On Figure 3, we have plotted the different modified energy ratios for three levels of noise. On the basis of these plots, we decided Han, Wong, and Bancroft 4 CREWES Research Report — Volume 21 (2009) to use er3(i) version for general time-picking. In the rest of this report, the MER acronym refers to er3(i). Comparison of the STA/LTA and MER time picking methods On Figure 4, we have plotted the STA/LTA ratio and MER for low-noise and highnoise signals. For high SNR arrivals, both techniques yield arrival times very close to the first-break time. For signals with low SNR, the picks for both methods occur somewhat later and closer to a peak absolute amplitude. FIG. 4: STA/LTA and MER time-picking on a clean trace and on a noisy trace. The noisy trace (SNR=3) was created by adding synthetic random noise to the clean trace (SNR=100). For the clean trace, the MER and STA/LTA methods both give an arrival time at the first break. For the noisy trace, both the STA/LTA and MER picks occur at a later time. Testing on field seismograms Figure 5 shows time picks from field seismograms recorded with an array of twelve 3C geophones spaced 11.75m apart in an observation well located about 575m from a treatment well. The seismograms were generated by a casing perforation shot at a depth of 2150m in the treatment well. Synthetic random noise was added to the raw seismograms to yield the plotted traces (SNR approximately 3). We see that the picks using the MER method appear to be more consistent and accurate than the STA/LTA picks. In addition, the MER method was almost five times faster than the STA/LTA method (this is due to the fact that LTA energy window is much longer than the MER energy windows). Testing on synthetic seismograms Gathers of synthetic microseismograms were created to represent recordings by a microseismic monitoring system. The system consisted of an array of twelve 3C geophones spaced vertically at 25m in an observation well. The distance between the microseismic source and the geophone array is assumed to be 500m. Microseismic noise reduction and time picking CREWES Research Report — Volume 21 (2009) 5 FIG. 5: MER and STA/LTA time picks for noisy field microseismograms. The vertical component traces for the geophones are shown in red, while the horizontal component traces are shown in blue. The red triangles are the first arrival time picks. The time picks from the MER method (top) are more consistent and accurate than those from the STA/LTA method (bottom). Han, Wong, and Bancroft 6 CREWES Research Report — Volume 21 (2009) The source signal is simulated by a minimum-phase wavelet of the form ww(tt) = sin⁡(2ππππtt) · exp⁡(−kktt) , where t is in seconds, f is 80Hz or 200Hz, and k is 50 nepers/s. The earth model used to calculate arrival times through ray-tracing is homogeneous and isotropic in P-wave or Swave velocities. A particular geophone signal is synthesized by convolving the source wavelet with a delta function located at the calculated arrival time. Various levels of Gaussian random noise are then added in order to study the limitations on the time picking methods caused by noise. FIG. 6: Time picking using the ER (top) and the MER (bottom) attributes on synthetic seismograms with Gaussian noise (SNR=2.5). The ER attribute picks incorrect arrival times for many traces. The MER attribute picks correctly for all traces. Microseismic noise reduction and time picking CREWES Research Report — Volume 21 (2009) 7 Figure 6 shows why the MER attribute should be used instead of the basic energy ratio ER for time picking when SNRs are low. The ER picks are incorrect for a good number of the traces in the synthetic gather. On the other hand, the MER picks are all correctly placed near the first break times of the arrivals. Trace windowing for traces with extremely low SNRs When traces are extremely noisy, the MER time picks often have outliers. This is the case for the picks on Figure 7, where five of the MER picks (plotted in green) are outliers. Applying a 5-point median filter changes the picks to the green crosses and eliminates the worst outliers. However, the filtered picks are incorrect for many of the traces. We must re-do the picking, but only after windowing the traces. FIG. 7: Left: the red crosses are initial MER time picks, showing five outliers (circled in red); the green crosses are the picks after 5-point median filtering. Right: MER picks after applying a window (defined following the steps in the text) to eliminate the outliers. Define the window in the following way. Let tmin and tmax be the m