Autocorrelation: Definition, Seismic Deconvolution, and Wavelets
Autocorrelation is a mathematical operation that measures the similarity of a signal with a time-shifted version of itself. In quantitative terms, the autocorrelation function R(τ) is defined as the integral of the product of a function f(t) and its delayed copy f(t+τ) over all time: R(τ) = ∫f(t)f(t+τ)dt. For discrete seismic data sampled at interval Δt, the summation form R(k) = Σx(n)x(n+k) is applied over all sample indices n. The result is a symmetric function of lag τ (or lag index k) that reveals the periodic structure, embedded wavelets, and noise character of the original signal without requiring any external reference trace.
In seismic exploration and reservoir characterization, autocorrelation is indispensable for wavelet estimation, deconvolution filter design, multiple identification, and spectral analysis. Because recorded seismic traces are the convolution of the earth's reflectivity series with the seismic wavelet plus noise, the autocorrelation of those traces encodes the wavelet's autocorrelation. Exploiting that relationship is the conceptual foundation of predictive deconvolution and minimum-phase Wiener filter design, two of the most widely used seismic processing steps employed worldwide from the offshore platforms of the Norwegian Continental Shelf to the tight-sand plays of the Permian Basin.
Key Takeaways
- The autocorrelation of a seismic trace is symmetric about zero lag and reaches its absolute maximum at zero lag, where its value equals the total signal energy.
- Normalizing the autocorrelation to unity at zero lag produces the autocorrelation coefficient, which ranges from -1 to +1 and allows direct comparison across traces with different amplitude scales.
- Periodic events such as water-bottom multiples produce distinct side-lobes in the autocorrelation at lags equal to the two-way travel time of the water column, making the autocorrelation a reliable multiple diagnostic tool.
- The power spectral density of a stationary random process equals the Fourier transform of its autocorrelation function, a relationship known as the Wiener-Khinchin theorem, which connects time-domain autocorrelation analysis directly to frequency-domain spectral whitening and bandwidth estimation.
- Autocorrelation should not be confused with cross-correlation: in vibroseis acquisition the pilot sweep is cross-correlated with the raw record to compress the long sweep into an impulsive response, while autocorrelation is used subsequently in processing to estimate the embedded wavelet and design deconvolution operators.
How Autocorrelation Works in Seismic Processing
The fundamental properties of the autocorrelation function make it a practical tool in seismic data processing. First, it is an even function: R(τ) = R(-τ), so the function is perfectly symmetric about the zero-lag axis. This means a processor only needs to compute and display the positive-lag side. Second, the zero-lag value R(0) equals the sum of squared amplitudes of the original signal, which is the signal's total energy. All other lags satisfy |R(τ)| ≤ R(0), so the function always peaks at zero lag. Third, the normalized autocorrelation, obtained by dividing every lag value by R(0), produces a coefficient that ranges from -1 to +1, where +1 at zero lag is guaranteed. These three properties allow geophysicists to quickly assess signal quality: a broad, slowly decaying normalized autocorrelation indicates a long, ringy wavelet with narrow bandwidth; a sharp spike at zero lag with rapid decay to near-zero at adjacent lags indicates a compact, broadband wavelet close to a minimum-phase spike.
In the Wiener filter framework underpinning predictive deconvolution, the autocorrelation matrix of the seismic trace (often called the Toeplitz matrix) appears on the left-hand side of the normal equations. The filter designer specifies a prediction lag equal to the expected wavelet length and solves for a filter that predicts, and therefore subtracts, the predictable portion of the trace. What remains after subtraction is the unpredictable component, ideally an approximation to the earth's reflectivity. Computing the autocorrelation over a statistically representative window (typically 500 ms to 2,000 ms in length) is the first computational step. The quality of the deconvolution result is directly tied to how accurately that window captures the wavelet's autocorrelation, which is why processors examine autocorrelation panels across many traces and multiple windows before committing to a single set of deconvolution parameters.
Spectral analysis connects directly to autocorrelation through the Wiener-Khinchin theorem: the power spectral density (PSD) of a wide-sense stationary process equals the Fourier transform of its autocorrelation function. In practice, this means that computing the discrete Fourier transform of the autocorrelation panel gives a smoothed estimate of the amplitude spectrum squared. Processors use this relationship to estimate bandwidth, identify notch frequencies introduced by near-surface ghosting or receiver arrays, and design spectral equalization (whitening) operators. If the autocorrelation PSD shows a strong notch at a particular frequency, the deconvolution filter will attempt to boost energy there, which can amplify noise if the signal-to-noise ratio at that frequency is poor. Recognizing this trade-off is essential for producing reliable seismic attribute volumes and amplitude maps downstream in the workflow.
Multiple Identification and Noise Characterization
One of the most operationally valuable uses of autocorrelation in seismic exploration is the identification of periodic multiples. A water-bottom multiple arrives at a two-way time equal to twice the water-column travel time after the primary reflection. If the water depth is 250 m (820 ft) and the water velocity is 1,500 m/s (4,921 ft/s), the water-bottom multiple period is approximately 333 ms. In the autocorrelation of any trace that contains this multiple, a pronounced positive side-lobe appears at lag 333 ms, and additional side-lobes appear at integer multiples of 333 ms. The autocorrelation display therefore serves as a diagnostic: if the interpreter observes evenly spaced side-lobes, they can immediately estimate the water depth, quantify the multiple period with sub-millisecond precision, and design surface-related multiple elimination (SRME) or predictive deconvolution operators tuned to that period. In shallow-water marine surveys such as those common in the North Sea or offshore West Africa, this diagnostic step is performed routinely on raw field data before any other processing.
Noise characterization is equally important. White (random) noise has a flat power spectrum, which by the Wiener-Khinchin theorem means its autocorrelation is a spike at zero lag and zero at all other lags. In reality, seismic noise is rarely truly white: coherent noise sources such as swell noise, cable strum, or 60 Hz (50 Hz in many international jurisdictions) power-line interference produce non-zero off-lag peaks at specific frequencies. By examining the off-lag character of the autocorrelation, a processor can quantify the fraction of coherent noise in the record, identify its periodicity, and design targeted noise-suppression filters. This is particularly relevant in land seismic acquisition in regions with dense infrastructure, such as the Alberta foothills in Canada or the Permian Basin in West Texas, where power-line and cultural noise routinely contaminate field records. Comparing autocorrelation panels computed on pre- and post-noise-attenuation datasets is a standard quality-control step used to verify that the noise filter removed coherent energy without harming the primary signal.
Wavelet Estimation and Deconvolution Design
The seismic convolutional model states that a recorded trace x(t) equals the convolution of the earth's reflectivity r(t) with the seismic wavelet w(t), plus noise n(t): x(t) = r(t) * w(t) + n(t). If the reflectivity is assumed to be white (uncorrelated), the autocorrelation of the recorded trace equals the autocorrelation of the wavelet alone (plus a noise term at zero lag). This is the statistical wavelet estimation premise: by computing the autocorrelation of a representative trace or ensemble of traces, the processor obtains the autocorrelation of the embedded wavelet. The wavelet itself can then be recovered by spectral factorization (extracting the minimum-phase wavelet consistent with the observed autocorrelation) or by constrained optimization if a well tie is available to provide a deterministic phase estimate.
In well-tie workflows, the synthetic seismogram generated by convolving the acoustic impedance log with a wavelet extracted from the autocorrelation of nearby seismic data is compared to the actual seismic trace at the well location. A good match validates the wavelet estimate and the acoustic log calibration simultaneously. Poor matches often indicate that the minimum-phase assumption used in spectral factorization is incorrect, prompting the processor to introduce a phase rotation derived from the well tie. This iterative loop between autocorrelation-based wavelet estimation, well-tie assessment, and phase correction is central to reservoir characterization workflows in basins as geologically diverse as the Permian Basin, the Norwegian Continental Shelf, the Cooper Basin in Australia, and the giant carbonate fields of the Middle East.