Space-Frequency Domain

Key Takeaways

• The predictability of coherent seismic events in the f-x domain is the mathematical key that separates signal from noise in the most effective random noise attenuation algorithms — a coherent seismic reflection event (whether flat, dipping, or curved) occupies a predictable relationship between spatial position and phase at each temporal frequency; in the f-x domain, a plane-wave event appears as a complex sinusoid in x at each frequency, meaning that the value at one spatial location can be predicted from the values at neighboring locations using a short finite-impulse-response (FIR) spatial prediction filter; random noise, by contrast, has no spatial predictability (its phase is random at each location and each frequency) and therefore cannot be predicted by the filter; when the prediction filter is applied to f-x data, it produces a "predicted" version of the data that contains only the coherent, spatially predictable events and lacks the random noise; subtracting the original data from the prediction error (or equivalently, using the prediction filter output as the noise-attenuated data) removes the random noise while preserving the coherent reflections; the effectiveness of f-x prediction depends on the assumption that coherent events are indeed spatially predictable (which fails for steeply dipping events at high frequencies or for spatially aliased data), and the choice of filter length (in spatial samples) determines the tradeoff between effective noise suppression and potential smearing of lateral reflection amplitude variations that could indicate DHI effects.
• The f-k domain (frequency-wavenumber domain, obtained by transforming seismic data in both time and space) is a related but distinct representation used for a different class of signal and noise separation problems — in the f-k domain, both the time and spatial axes are Fourier transformed, so the data is represented as a function of temporal frequency (f, in Hz) and spatial frequency or wavenumber (k, in cycles per meter); coherent seismic events appear in the f-k domain as bands of energy concentrated around the apparent dip velocity of the event, because a plane-wave event at dip velocity v occupies the line f = v x k in the f-k plane; this concentration allows coherent noise that has a different apparent velocity from the signal (such as ground roll in land seismic, which travels at 200-800 m/s while primary reflections appear at 1,500-5,000 m/s) to be separated from the signal by designing a fan-shaped rejection filter in the f-k plane that removes energy below a velocity threshold; f-k filtering is particularly effective for coherent linear noise suppression but can damage signal energy when signal and noise velocities overlap, which is why f-x and f-k methods are applied to different noise types and used complementarily in the processing workflow.
• F-x-y deconvolution for 3D seismic data extends the 2D f-x concept to three spatial dimensions, providing noise suppression that respects the volumetric coherence of 3D reflection events — in 3D seismic processing, f-x-y deconvolution applies the spatial prediction filter concept simultaneously in both the in-line (x) and cross-line (y) spatial directions at each temporal frequency, exploiting the spatial coherence of 3D reflection events in both lateral directions; this 3D approach is more effective than applying 2D f-x deconvolution independently on successive in-line and cross-line panels, because 3D events have true volumetric coherence that 2D filters underutilize; the computational cost of f-x-y deconvolution is substantially higher than 2D f-x processing, requiring more sophisticated algorithms (patch-based processing, Wiener filter solutions in 3D) and more computing resources; the improvement in data quality from 3D versus 2D f-x processing is most pronounced in datasets with strongly 3D noise (such as acquisition footprint noise from irregular receiver geometry or from ocean-bottom cable gaps) where 2D filters operating on individual in-line or cross-line panels cannot distinguish noise from signal without the additional spatial dimension.
• Amplitude-versus-offset (AVO) analysis benefits from f-x domain processing that preserves amplitude fidelity while suppressing noise — AVO analysis measures how seismic reflection amplitude varies with source-receiver offset (angle of incidence), using the offset-dependence as an indicator of reservoir fluid type (gas-saturated sands typically show amplitude increase with offset, while brine-saturated sands show amplitude decrease or no change); this amplitude analysis is compromised by random noise, which creates apparent AVO anomalies in noisy data that can be mistaken for genuine fluid indicators; f-x deconvolution is widely used as a pre-AVO conditioning step because it suppresses random noise without distorting the reflection amplitude as a function of offset (when applied within offset-limited ranges or after offset regularization); alternative amplitude-preserving random noise attenuation methods (rank-reduction algorithms, t-x prediction error filtering, curvelet transforms) are evaluated specifically on their ability to suppress noise without attenuating or creating spurious AVO gradients, because a false AVO anomaly generated by processing is as damaging to exploration decisions as a real anomaly masked by noise.
• Machine learning approaches to seismic noise attenuation are being benchmarked against f-x deconvolution as the reference classical method — the f-x domain prediction filter represents a well-understood, theoretically grounded approach to seismic signal and noise separation that has been refined over 40 years of industrial application; its performance and failure modes are well characterized: it works well for random noise on regular-geometry acquisition data with moderately dipping events, and it fails for highly irregular geometry, steeply dipping events, and non-stationary noise; machine learning denoising algorithms (convolutional neural networks, U-Net architectures trained on synthetic or field data) are showing promising results that exceed classical f-x methods in some respects (better handling of non-stationary noise, ability to remove coherent noise types that resist f-x prediction) while struggling in others (generalization to acquisition geometries and noise types not well represented in training data); the comparison between ML and f-x methods in published seismic processing benchmarks consistently shows that f-x deconvolution remains competitive with ML in its core application domain of random noise suppression on regular-geometry land and marine surveys, while ML methods have a potential advantage in irregular-geometry data and in suppressing coherent acquisition noise patterns that are difficult to model analytically.