Convolution: Seismic Trace Modeling and Deconvolution

What Is Convolution?

Convolution (also called the convolutional model of the seismic trace) is the mathematical operation that describes how a recorded seismic trace is produced as the convolution of the earth's reflectivity series with the seismic wavelet, expressed as Trace = Reflectivity * Wavelet + Noise. Each interface in the subsurface generates a reflection coefficient spike, and the recorded waveform at that time is a scaled copy of the source wavelet; the complete trace is the sum of all such overlapping wavelets. Deconvolution reverses this process to recover the reflectivity series, which represents the true geological contrasts stripped of the source signature.

Key Takeaways

The convolutional model states: s(t) = r(t) * w(t) + n(t), where s(t) is the seismic trace, r(t) is the reflectivity series, w(t) is the wavelet, and n(t) is noise.
Convolution in the time domain is equivalent to multiplication in the Fourier frequency domain, enabling fast computation via the FFT algorithm.
Spiking (whitening) deconvolution compresses the wavelet toward a spike by flattening the amplitude spectrum, enhancing temporal resolution.
The convolutional model assumes a minimum-phase, stationary wavelet and no multiples — simplifications that require corrections in practice.
Predictive deconvolution attenuates short-period multiples by predicting and subtracting periodic energy at a lag equal to the multiple's period.

The Mathematics of Convolution in Seismic Processing

Formally, convolution of two time functions f(t) and g(t) is defined as the integral of f(tau) times g(t minus tau) over all tau — the operation slides one function across a time-reversed copy of the other, computing their overlap at every time shift. In discrete form used in digital seismic processing, this becomes a sum of products between the samples of one series and the reversed samples of the other, shifted one sample at a time. The output is longer than either input by (N + M - 1) samples, which explains why convolution with a wavelet of finite length spreads each reflectivity spike into a waveform lasting the duration of the wavelet. This spreading is exactly the temporal blurring that limits the ability to resolve thin beds whose thickness is less than approximately one-quarter of the dominant wavelength.

The equivalence between time-domain convolution and frequency-domain multiplication is the cornerstone of efficient seismic processing. A 4,000-sample trace convolved with a 200-sample wavelet requires roughly 800,000 multiplications in the time domain but only 4,200 complex multiplications using the FFT — a speedup of more than 100x. In practice, the entire processing flow from noise attenuation to migration exploits frequency-domain implementations. Deconvolution fits naturally in this framework: the deconvolution operator is the inverse filter whose frequency-domain spectrum is the reciprocal of the estimated wavelet spectrum, modified by a white-noise stabilization factor (the water level) to prevent division by near-zero values at notch frequencies.

Wavelet estimation is the critical unsolved problem in deconvolution. Statistical methods assume the earth's reflectivity is white (all reflection coefficients equally probable at all frequencies) so that the autocorrelation of the seismic trace equals the autocorrelation of the wavelet alone. Under the additional assumption of minimum phase, the wavelet can be extracted from the trace autocorrelation without a well. Deterministic wavelet extraction uses a nearby well log to compute a synthetic seismogram, then cross-correlates it with the trace to extract the wavelet directly — this is more accurate but requires a well with good sonic and density logs within the survey area. Neither method is exact, and residual wavelet error is a primary cause of misties between wells and seismic in prospect evaluation.

Fast Facts: Convolution

Convolutional model equation: s(t) = r(t) * w(t) + n(t); asterisk denotes convolution
Frequency domain equivalent: S(f) = R(f) x W(f) + N(f); multiplication replaces convolution
Thin bed resolution limit: approximately lambda/4, where lambda is dominant wavelength (e.g., 10 m at 50 Hz in 2,000 m/s rock)
Typical seismic wavelet duration: 40-100 ms for surface seismic; 20-40 ms for VSP
Spiking deconvolution output: broadband reflectivity spike train with white amplitude spectrum within the signal bandwidth
Predictive deconvolution lag: set equal to primary-to-multiple delay, typically 50-200 ms for water-bottom multiples
Water-level stabilization: 0.1-1% of the peak power spectrum added before inversion to prevent instability
Minimum phase assumption: energy front-loaded in time; not always valid for vibroseis (zero phase) sources requiring phase rotation

Processing Tip:

Before running spiking deconvolution on a dataset, examine the average autocorrelation of a long window of raw traces. If the autocorrelation shows prominent side lobes at regular lags (50 ms, 100 ms, 150 ms), these are water-bottom multiples; switch to predictive deconvolution with the lag set to the dominant side-lobe position. Applying spiking deconvolution to a multiple-contaminated dataset will whiten the multiple energy along with the primary reflectivity, producing a result that looks sharp but contains residual multiples that mislead the interpreter.

Convolution is also referred to as:

Linear filtering — convolution with a filter operator is the mathematical definition of linear filtering in signal processing
Superposition integral — the integral form of convolution reflects the principle of superposition in linear systems
Convolutional model — shorthand for the full model s(t) = r(t) * w(t) + n(t) used throughout seismic interpretation literature
Fold — the German-origin term (Faltung) for convolution appears in some mathematical and geophysics texts

Frequently Asked Questions About Convolution

Why does deconvolution improve seismic resolution?

A recorded seismic trace has a band-limited amplitude spectrum shaped by the source wavelet, which typically peaks around 20-60 Hz and rolls off at both low and high frequencies. Deconvolution applies an inverse filter that amplifies the frequencies where the wavelet is weak, flattening the spectrum toward white noise within the usable signal bandwidth. This whitening in the frequency domain corresponds to wavelet compression toward a spike in the time domain, so individual reflection events that were blurred over the wavelet duration of 50-80 ms become sharper, allowing geologists to distinguish thin beds that were previously unresolvable.

What happens when the minimum-phase assumption is violated?

Vibroseis sources produce a zero-phase Klauder wavelet after correlation, not a minimum-phase wavelet. If the standard autocorrelation-based deconvolution operator is applied without first accounting for the source phase, the output will have a residual phase error that shifts reflection events in time and distorts their shape. The standard remedy is to estimate the source wavelet deterministically from a well tie and apply a phase rotation operator (or a full shaping filter) to convert the data to a true zero-phase or minimum-phase state before spiking deconvolution. Unrecognized phase errors are a common cause of poor well-to-seismic ties that persist even after careful depth conversion.

How does the stationary wavelet assumption break down in practice?

The convolutional model assumes the wavelet shape is constant with depth (time), but in reality high-frequency energy is absorbed preferentially as waves travel through rock (a process called intrinsic attenuation, quantified by the quality factor Q). At depths of 3,000-5,000 m, the dominant frequency may have shifted from 50 Hz near surface to 25-30 Hz at the target, and the wavelet phase may have rotated significantly. Processing workflows address this with Q-compensation (inverse-Q filtering) before deconvolution, or with time-variant deconvolution that applies different operator designs in successive time windows. Ignoring the non-stationary wavelet degrades resolution at depth while over-sharpening shallow reflectors.

Why Convolution Matters in Oil and Gas

The convolutional model is the theoretical foundation on which every seismic interpretation workflow rests. Understanding that a seismic trace is a blurred, noise-contaminated version of the earth's reflectivity — and that deconvolution is the tool to undo that blurring — is essential for correctly interpreting thin reservoir sands, evaluating DHI anomalies, and building reliable synthetic seismograms for well ties. Errors in wavelet estimation propagate directly into AVO analysis, acoustic impedance inversion, and depth conversion, affecting every subsequent step from prospect mapping to reservoir characterization. For geophysicists working in complex basins with strong multiples, gas-driven attenuation, or mixed-phase sources, a rigorous grasp of the convolutional model separates reliable interpretations from artifacts that can misdirect drilling decisions worth tens of millions of dollars.