Random Noise

Key Takeaways

• CMP stacking is the primary and most robust method of random noise attenuation because it exploits the statistical independence of random noise between traces while coherently summing the reflected signal: in a CMP gather with 60-fold coverage (60 traces sharing the same midpoint), the reflected signal from a target horizon adds coherently (after normal moveout correction to flatten the reflection hyperbola) to produce a stacked amplitude equal to 60 times the single-trace signal amplitude, while random noise from 60 independent traces adds incoherently to produce a stacked noise amplitude equal to the square root of 60 (approximately 7.7) times the single-trace noise amplitude; the resulting SNR improvement from stacking 60 traces is approximately 60/7.7 = 7.7 (the square root of 60), compared to a single unstacked trace; in very noisy environments (deep water with swell noise, land areas with strong wind or cultural noise), higher fold acquisition (120-fold, 240-fold, or even higher for ultra-high-resolution surveys) is specified precisely to achieve the SNR improvement needed to image the target horizons above the noise floor; the fold must be earned through appropriate source and receiver spacing in the acquisition geometry, and the stacking effectiveness depends on the NMO correction being accurate enough to align all reflections before stack — velocity errors that produce residual moveout in the CMP gather cause partial destructive interference of the signal during stack, reducing the realized SNR improvement below the theoretical square-root-of-fold maximum.
• Random noise attenuation in the processing flow uses a suite of algorithms applied before and after stacking to suppress noise that stacking alone cannot eliminate: f-x deconvolution (also called f-x prediction) is the most widely used post-stack random noise filter, operating in the frequency-space (f-x) domain where a predictable seismic signal is modeled as a sum of complex exponentials (plane waves) and the residual (unpredicted) energy is treated as noise; the filter predicts the seismic amplitude at each trace from its neighbors using a Wiener-Levinson prediction filter, and the unpredictable residual (random noise) is subtracted from the data; f-x deconvolution is effective when reflections are laterally continuous (as in deep-water sedimentary sequences with good inter-trace correlation), but it suppresses legitimate lateral amplitude variations that may indicate lithology or fluid changes, requiring conservative application that preserves DHI-quality amplitude information; 3D median filtering, Karhunen-Loeve (K-L or eigenimage) filtering, and structure-oriented filtering are additional random noise suppression algorithms used in specific situations; all noise attenuation algorithms involve a tradeoff between noise suppression and signal fidelity, and over-processing that removes noise along with genuine stratigraphic signal is a significant risk in exploration seismic processing that requires careful QC against input data.
• Ambient noise recording and noise floor characterization before a seismic survey is conducted as part of survey design to ensure that the acquisition parameters (source energy, sweep length for vibroseis, number of sweeps, recording time, geophone type) are appropriate for the noise environment at the survey location: the noise floor measurement involves deploying geophones at the survey location without active seismic sources and recording the ambient ground motion for several minutes, then analyzing the frequency spectrum of the recorded noise to identify the dominant noise frequencies (typically low frequency for wind and traffic, high frequency for electronic noise) and amplitudes; the source energy and recording parameters are then designed so that the desired reflection signal level exceeds the ambient noise floor by a sufficient margin (typically at least 6 dB, preferably 20 dB) at the target frequency band; in areas with severe ambient noise (near cities, highways, wind farms, or industrial facilities), the noise floor may be so high that the seismic source cannot achieve adequate SNR even with many sweeps or a powerful explosive source, and the survey may need to be conducted during low-noise periods (nights, weekends, low-wind periods) or noise cancellation methods must be used; in marine acquisition, the ambient noise floor is dominated by sea state and vessel noise, and rough weather conditions (high sea states) may cause the survey to be suspended when noise levels exceed a threshold that would degrade data quality below acceptable limits.
• The distinction between random noise and coherent noise determines which processing methods are appropriate for noise suppression and which will inadvertently suppress the signal along with the noise: random noise has no preferred moveout direction and cannot be separated from signal based on moveout alone, requiring statistical suppression methods (stacking, f-x prediction) that rely on the signal being more predictable than the noise; coherent noise (surface waves, guided waves, air waves, multiples) has specific moveout velocities and can be separated from primary reflections by velocity-based filtering (f-k filtering, radon transform muting, tau-p filtering) that removes energy at noise velocities while preserving energy at signal velocities; misidentification of coherent noise as random noise (and vice versa) leads to inappropriate processing choices that either fail to remove the coherent noise (if stacking-based methods are used against moveout-coherent noise) or fail to preserve the signal while removing coherent noise (if velocity-based filters are applied with overly broad muting windows that include signal energy); the seismic processor's skill in distinguishing noise types from the raw shot records, examining the frequency content, moveout, and spatial patterns of the recorded energy, determines the quality of the noise attenuation workflow and ultimately the quality of the processed seismic volume used for reservoir characterization.
• Random noise in seismic interpretation (as opposed to data acquisition and processing) refers to the background of non-reflective energy present on processed seismic sections and volumes that makes the picking of weak or discontinuous reflections uncertain: even after optimal processing, a residual random noise level remains on seismic sections, and the detectability of a reflection depends on the local signal-to-noise ratio at the reflection depth — a reflection that is clearly visible (high SNR) at shallow depths may become uncertain or invisible at deeper depths where geometric spreading, anelastic attenuation, and multiple contamination have all degraded the signal; automated horizon tracking algorithms that pick reflections across a 3D seismic volume are sensitive to random noise levels because they can follow noise spuriously picked as a weak reflector, creating artificial artifacts in the picked surface that must be identified and corrected by the interpreter; in thin-bed tuning studies where the reservoir is below the quarter-wavelength tuning thickness and the reflection amplitude must be extracted with sub-sample precision, the random noise level on the seismic data directly limits the resolution of the amplitude extraction and the reliability of the porosity or fluid saturation estimate derived from the amplitude; quantifying the noise level on seismic data (by measuring trace-to-trace amplitude variance in geologically simple areas where no lateral reflection amplitude variation is expected) is a basic QC step in any seismic reservoir characterization study.