Skip to content

Seismic Data Processing: CDP, Velocity Analysis, F-K Filtering, Stacking, Deconvolution and Migration

CSIR NET seismic processing: convolutional model, statics, CMP gathers, semblance, NMO and stacking, Wiener deconvolution, f-k filtering, migration.

  • 19 min read
  • Updated
On this page · 17 sections
  1. Prerequisites
  2. The Hook: From Mess to Section
  3. The Convolutional Model
  4. The Standard Processing Flow
  5. CDP, CMP and Gathers
  6. Velocity Analysis
  7. Stacking
  8. Deconvolution
  9. Frequency and F-K Filtering
  10. Migration
  11. Memory Tricks
  12. Don’t Confuse
  13. Common Mistakes
  14. Practice
  15. Quick Revision Table
  16. Key Takeaways
  17. What’s Next

Prerequisites

The Hook: From Mess to Section

How CSIR NET asks it. Part C (Geophysics) asks for the logic of the flow (why NMO before stack, why decon before stack), quick numbers (fold, \(\sqrt N\), aliasing in f-k, static shifts, migrated dip and lateral shift, NMO stretch), small deconvolution calculations with two-term wavelets, and statement-type questions on the assumptions of spiking deconvolution and the behaviour of migration with wrong velocities.


The Convolutional Model

A recorded trace is modelled as

$$\boxed{x(t) = w(t) * r(t) + n(t)} \qquad\Longleftrightarrow\qquad X(f) = W(f)\,R(f) + N(f)$$

where \(w\) is the wavelet (source signature × earth filter × recording system), \(r\) is the reflectivity series and \(n\) noise. Processing tries to remove \(w\) (deconvolution), suppress \(n\) (filtering, stacking) and put \(r\) in the right place (NMO, statics, migration).

Wavelet phase.

TypeEnergy distributionNotes
Zero-phaseSymmetric about \(t = 0\); non-causalBest resolution for a given amplitude spectrum; interpreters prefer it; Vibroseis Klauder wavelet
Minimum-phaseEnergy concentrated as early as possible; causalHas a causal, stable inverse; dynamite signatures are close to it; air-gun far-field signatures are often treated as approximately so
Maximum-phaseEnergy concentrated lateCausal inverse is unstable
Mixed-phaseIn betweenReal wavelets after filtering

In z-transform terms (\(z\) = unit delay), a two-term wavelet \((w_0, w_1)\) is minimum-phase when \(\lvert w_0\rvert > \lvert w_1\rvert\): the zero of \(w_0 + w_1z\) lies outside the unit circle.


The Standard Processing Flow

graph TD
    A["Field data:
demultiplex, format"] --> B["Geometry, trace edits"] B --> C["Gain recovery:
spherical divergence"] C --> D["Field statics:
elevation and weathering"] D --> E["Pre-stack deconvolution"] E --> F["CMP sort"] F --> G["Velocity analysis"] G --> H["NMO correction and mute"] H --> I["Residual statics,
velocity update"] I --> J["CMP stack"] J --> K["Post-stack filtering,
decon, noise attenuation"] K --> L["Migration"] L --> M["Display and interpretation"]

Gain and amplitude recovery

  • Spherical divergence correction multiplies by a function that grows with time (in a \(v(z)\) medium, roughly proportional to \(tV^2(t)\)) and an absorption term. It keeps relative amplitudes, which AVO and bright-spot work need.
  • Automatic gain control (AGC) scales each window by its own mean amplitude. It makes weak events visible but destroys relative amplitudes: never apply it before amplitude analysis.

Static corrections

Statics are time-invariant shifts per trace that remove the effects of surface elevation and the low-velocity weathered layer, referring all traces to a flat datum below which a replacement velocity applies.

  • Elevation (datum) static: \(-\Delta z/V_{rep}\) for each source and receiver above the datum.
  • Weathering static: replace travel time in the slow weathered layer by travel at \(V_{rep}\). With weathering thickness \(h_w\) and velocity \(V_w\), the correction to the base of weathering is \(-h_w/V_w\); refraction first breaks (uphole times, intercepts) estimate it.
  • Residual statics: small leftovers estimated after NMO by cross-correlating traces with a pilot, decomposed surface-consistently into source and receiver terms.

Uncorrected statics jitter reflections and lower stacking quality; long-wavelength statics errors create false structure.


CDP, CMP and Gathers

A common midpoint (CMP) gather collects traces whose source–receiver midpoints coincide. For flat layers, all of them reflect from one subsurface point, the common depth point (CDP). For dipping layers, reflection points spread up-dip of the midpoint (CMP smearing), so CMP and CDP are not the same; dip moveout (DMO) correction partly fixes this.

GatherFixedUse
Common-shotSourceField QC, noise analysis, pre-stack migration in shot domain
Common-receiverReceiverReceiver statics, receiver-domain filtering
Common-midpointMidpointVelocity analysis, NMO, stacking
Common-offsetOffsetNear-offset sections, AVO, pre-stack migration

Multiple coverage of the same midpoint (fold) is what makes velocity analysis and stacking possible.


Velocity Analysis

Velocity analysis finds, for each \(t_0\), the velocity that best flattens reflection hyperbolae on a CMP gather.

MethodHow it works
Constant-velocity stacks (CVS)Stack a panel of CMPs at many trial velocities; pick the velocity at which each event is sharpest
Constant-velocity gathers (CVG)NMO-correct one gather at trial velocities; pick where events are flat
Velocity spectrum (semblance)For each \(t_0\) and trial \(V\), measure coherence along the hyperbola; contour the result and pick peaks

Semblance for \(N\) traces with amplitudes \(a_{i,t}\) along a trial hyperbola, summed over a short time window:

$$S = \frac{\sum_t\left(\sum_{i=1}^{N}a_{i,t}\right)^2}{N\sum_t\sum_{i=1}^{N}a_{i,t}^2}, \qquad 0 \le S \le 1$$

\(S = 1\) when all traces are identical along the trial curve. Picking rules:

  • Primary velocities usually increase with time; picks that would give absurd Dix interval velocities are wrong.
  • Multiples appear as a lower-velocity trend (they travelled mostly in the shallow, slow section).
  • Velocity resolution worsens with depth, because moveout becomes small compared with the wavelet period. Long offsets help.

NMO correction, stretch and mute

NMO correction shifts each sample from \(t_x\) to \(t_0\) using the picked velocity. It stretches the wavelet, worst at early times and far offsets. The fractional stretch is

$$\frac{\Delta f}{f} \approx \frac{\Delta t_{NMO}}{t_0}$$

Example (\(V = 2000\) m/s, \(t_0 = 1.0\) s): at 1000 m offset, \(\Delta t_{NMO} = 118\) ms (about 12% stretch); at 2000 m, \(t_x = 1.414\) s and stretch is about 41%. Strongly stretched samples are muted (outer or stretch mute) before stacking; first breaks and ground roll are removed by front and surgical mutes.


Stacking

After NMO, the traces of a CMP gather are summed (and normalised by the number of live traces). Stacking:

  • improves the signal-to-random-noise ratio by \(\sqrt N\) for fold \(N\);
  • attenuates multiples, which are under-corrected by primary velocities and so are not flat (their residual moveout spreads them out on stacking);
  • attenuates other events with wrong moveout, such as some ground roll and out-of-plane energy;
  • reduces the data volume by the fold factor.

Variants: mean, weighted (down-weight noisy traces), median (resistant to spikes) and diversity stacking (weight by inverse local power).


Deconvolution

Aims: compress the wavelet towards a spike (improve vertical resolution), remove short-period reverberations and ghosts, and balance (whiten) the spectrum.

The inverse filter

If \(w(t)\) is known and minimum-phase, its inverse \(f(t)\) satisfies \(f * w = \delta\). For \(w = (1, a)\) with \(\lvert a\rvert < 1\):

$$\frac{1}{1 + az} = 1 - az + a^2z^2 - a^3z^3 + \cdots$$

an infinite but convergent series. Any truncation leaves an error, and if \(\lvert a\rvert > 1\) (maximum phase) the series diverges.

Wiener (least-squares) filters

Choose a finite filter \(f\) of length \(m\) to minimise the energy of the difference between the actual output \(f * x\) and a desired output \(d\). Setting derivatives to zero gives the normal equations:

$$\begin{bmatrix} r_0 & r_1 & \cdots & r_{m-1}\\ r_1 & r_0 & \cdots & r_{m-2}\\ \vdots & & \ddots & \vdots\\ r_{m-1} & \cdots & r_1 & r_0\end{bmatrix}\begin{bmatrix} f_0\\ f_1\\ \vdots\\ f_{m-1}\end{bmatrix} = \begin{bmatrix} g_0\\ g_1\\ \vdots\\ g_{m-1}\end{bmatrix}$$

where \(r_k\) is the autocorrelation of the input and \(g_k\) the cross-correlation of desired output with input. The matrix is Toeplitz (constant along diagonals), solved efficiently by the Levinson recursion.

VariantDesired outputPurpose
Spiking deconvolutionA spike at zero lagCompress the wavelet; whiten the spectrum
Predictive deconvolutionThe input itself, \(\alpha\) samples later (prediction lag/gap)Predict and subtract periodic events (reverberations, short-period multiples) while keeping the first \(\alpha\) of the wavelet
Shaping filterAny chosen waveletConvert to zero-phase, match surveys

Spiking deconvolution is predictive deconvolution with \(\alpha\) = one sample.

Assumptions of statistical (spiking) deconvolution:

  1. The reflectivity is white (random), so the trace autocorrelation has the shape of the wavelet autocorrelation.
  2. The wavelet is minimum-phase (the autocorrelation carries no phase information; minimum phase is assumed to recover it).
  3. The wavelet is stationary within the design window.
  4. Noise is small.

Practical parameters. Operator length (long enough to cover the wavelet or the reverberation period), design window, and prewhitening: adding a small percentage to \(r_0\) (white noise) stabilises the solution and keeps the filter from boosting noise at frequencies where the signal is weak.

Deterministic deconvolution uses a known wavelet: signature deconvolution of air-gun data with a measured far-field signature, or the zero-phase Klauder wavelet in Vibroseis data.

Water-layer reverberation

With a free surface (\(R = -1\)) and a sea-floor reflection coefficient \(R\), a spike generates a reverberation train with period equal to the water two-way time \(\tau\):

$$1 - 2Rz + 3R^2z^2 - 4R^3z^3 + \cdots = \frac{1}{(1 + Rz)^2}, \qquad z = \text{delay of } \tau$$

so the exact inverse is the three-point Backus filter \((1 + Rz)^2 = 1 + 2Rz + R^2z^2\), with coefficients at 0, \(\tau\) and \(2\tau\). Predictive deconvolution with a prediction lag equal to \(\tau\) achieves the same in practice.


Frequency and F-K Filtering

Band-pass filtering keeps the useful band (removing low-frequency ground roll and swell, high-frequency noise). Filters are usually designed zero-phase so that events do not shift.

F-K filtering. A 2D Fourier transform takes a gather from \((t, x)\) to frequency \(f\) and wavenumber \(k\). A linear event of apparent velocity \(V_a\) maps to a straight line through the origin:

$$V_a = \frac{f}{k}$$
  • High apparent velocity (near-vertical arrivals, flat reflections) plots close to the \(f\)-axis.
  • Slow events (ground roll, air wave, guided waves, steep linear noise) plot close to the \(k\)-axis.
  • A fan (pie-slice) filter rejects a wedge of slow apparent velocities and keeps the fast ones.
  • Events dipping the other way plot at negative \(k\), which is how VSP up-going and down-going waves are separated.
  • For multiples, apply NMO with a velocity between the primary and multiple trends: primaries are over-corrected (curve up), multiples under-corrected (curve down), so they dip in opposite directions and fall on opposite sides of the f-k plane. Today the \(\tau\)–\(p\) (Radon) transform is more often used for the same job.

Spatial aliasing. With trace spacing \(\Delta x\), the Nyquist wavenumber is

$$k_N = \frac{1}{2\Delta x}$$

An event of apparent velocity \(V_a\) is aliased above \(f = V_ak_N = V_a/(2\Delta x)\). Aliased energy wraps around to the opposite side of the f-k plane and can pass a fan filter. Slow noise on coarse spreads is almost always aliased, which is one reason field arrays are used.


Migration

What migration does

On an unmigrated (stacked, zero-offset) section each sample is plotted vertically below the midpoint, as if every reflection came from directly below. Dipping reflections actually come from up-dip. Migration:

  • moves dipping reflections up-dip;
  • makes them steeper and shorter;
  • collapses diffractions to their apex points;
  • unties bow-ties (buried foci) into synclines;
  • improves lateral resolution by shrinking the effective Fresnel zone.

Migrator’s equation (constant velocity, depth-scaled section):

$$\sin\theta_m = \tan\theta_u$$

where \(\theta_u\) is the apparent dip on the unmigrated section and \(\theta_m\) the true (migrated) dip. For time sections with measured time dip \(\Delta t/\Delta x\), velocity \(V\) and time \(t\):

$$\sin\theta_m = \frac{V}{2}\frac{\Delta t}{\Delta x}, \qquad d_x = \frac{V^2t}{4}\frac{\Delta t}{\Delta x}, \qquad t_m = t\sqrt{1 - \left(\frac{V\Delta t}{2\Delta x}\right)^2}$$

where \(d_x\) is the horizontal shift (up-dip) and \(t_m\) the migrated time.

Diffraction hyperbola from a point at depth \(z\) on a zero-offset section: \(t(x) = (2/V)\sqrt{z^2 + x^2}\), where \(x\) is the horizontal distance from the point. It is steeper than a CMP reflection hyperbola with the same apex time, \(t = (2/V)\sqrt{z^2 + x^2/4}\) with \(x\) the offset, because on a zero-offset section both source and receiver move away from the diffractor together.

Exploding-reflector model. A zero-offset section can be modelled as if every reflector exploded at \(t = 0\) and waves travelled one way to the surface at half the true velocity. This turns migration into one-way wavefield extrapolation.

Methods

MethodPrincipleStrengths and limits
Kirchhoff (diffraction summation)Sum amplitudes along each point’s diffraction hyperbola and place the sum at its apex, with obliquity and spreading weightsHandles steep dips; flexible geometry; aperture-limited
Finite-differenceDownward continuation of the wavefield with a finite-difference approximation of the (one-way) wave equationHandles vertical and lateral velocity variation; steep-dip accuracy depends on the approximation order
F-K (Stolt)Mapping in the Fourier domainVery fast; exact for constant velocity only (stretch tricks for \(v(z)\))
Phase-shift (Gazdag)Downward continuation by phase shifts, layer by layer in \(f\)-\(k\)Exact for \(v(z)\); no lateral variation
Reverse-time migrationFull two-way wave equationHandles complex structures, turning waves; costly

Time vs depth, post-stack vs pre-stack.

  • Time migration assumes a locally smooth velocity; it fails under strong lateral velocity changes (salt, overthrusts). Depth migration uses an interval-velocity model and handles them.
  • Post-stack migration assumes the stack is a good zero-offset section. Pre-stack migration handles conflicting dips and complex velocity, and pre-stack depth migration (PSDM) is the standard for complex structure; its imaging gathers also drive velocity-model building.

Velocity errors.

  • Velocity too low: under-migration; diffractions are not fully collapsed (residual “frowns”).
  • Velocity too high: over-migration; events turn into upward-concave “smiles”.
  • Noise bursts and data edges also produce smiles; tapers and apertures control them.

Memory Tricks

Don’t Confuse

ThisvsThat
CMP: common source–receiver midpointvsCDP: common subsurface reflection point; same only for flat layers
Spiking deconvolution: desired output a spikevsPredictive deconvolution: desired output the input \(\alpha\) later; removes periodic multiples
Autocorrelation (left side of normal equations)vsCross-correlation of desired output with input (right side)
Stacking velocity: flattens a CMP gathervsMigration velocity: images the section; different under dip and lateral change
Time migration: smooth lateral velocityvsDepth migration: strong lateral velocity variation
Under-migration: frowns remain, V too lowvsOver-migration: smiles, V too high
NMO: offset correction, gather domainvsDMO: dip correction that removes CMP smearing before stack
Statics: same shift for the whole tracevsNMO: time-variant shift

Common Mistakes


Practice

All questions below are original practice questions written in CSIR NET Part C (Geophysics) style. They are not previous-year questions.

Level 1: Foundation

  1. Name two assumptions of spiking (statistical) deconvolution.
  2. On an f-k plot, what quantity is given by \(f/k\) for a linear event?
  3. In which direction does migration move a dipping reflection?
Answers
  1. Any two of: white reflectivity, minimum-phase wavelet, stationarity, low noise.
  2. The apparent velocity of the event across the traces.
  3. Up-dip (and the event becomes steeper and shorter).

Level 2: Part B Style

Q1 (practice). Traces are spaced 25 m apart. Ground roll has an apparent velocity of 500 m/s. Above what frequency is it spatially aliased?

  • (a) 5 Hz
  • (b) 10 Hz
  • (c) 20 Hz
  • (d) 40 Hz
Answer and explanation

(b). \(k_N = 1/(2\Delta x) = 0.02\) cycles/m. Aliasing begins at \(f = V_ak_N = 500\times0.02 = 10\) Hz. Most ground-roll energy above 10 Hz will wrap around in the f-k plane.

Q2 (practice). A receiver sits on a 10 m weathered layer of velocity 500 m/s. The datum is 20 m below the surface and the replacement velocity is 2000 m/s. The static correction for this receiver is:

  • (a) −20 ms
  • (b) −25 ms
  • (c) −15 ms
  • (d) −5 ms
Answer and explanation

(b). Remove 10 m of weathering (10/500 = 20 ms) and 10 m of sub-weathering material above the datum at the replacement velocity (10/2000 = 5 ms): total −25 ms. Option (a) stops at the base of weathering.

Q3 (practice). On a zero-offset stacked section with velocity 2000 m/s, a reflection at 2.0 s has time dip 0.4 s/km. After migration its true dip and horizontal shift are about:

  • (a) 21.8° and 800 m down-dip
  • (b) 23.6° and 800 m up-dip
  • (c) 23.6° and 400 m up-dip
  • (d) 21.8° and 400 m up-dip
Answer and explanation

(b). \(\sin\theta_m = (V/2)(\Delta t/\Delta x) = 1000\times0.0004 = 0.4\) → \(\theta_m = 23.6°\). The unmigrated apparent dip is \(\tan^{-1}0.4 = 21.8°\) (the distractor). \(d_x = V^2t(\Delta t/\Delta x)/4 = (4\times10^6)(2.0)(4\times10^{-4})/4 = 800\) m, up-dip. Migrated time \(= 2.0\sqrt{1 - 0.16} = 1.83\) s.

Q4 (practice). Which processing step should be avoided before an amplitude-versus-offset study?

  • (a) Spherical-divergence correction
  • (b) Short-window automatic gain control
  • (c) Surface-consistent residual statics
  • (d) NMO correction with a stretch mute
Answer and explanation

(b). AGC rescales each window by its own amplitude and erases the relative amplitude changes with offset that AVO measures. Spherical-divergence correction is deterministic and amplitude-preserving.

Level 3: Part C Style

Q1 (practice). A wavelet has samples \(w = (1, 0.5)\).

(i) Write the first four terms of its exact inverse. (ii) Convolve the two-term truncated inverse \((1, -0.5)\) with \(w\) and find the error energy relative to a unit spike. (iii) Design the two-term least-squares (Wiener) spiking filter and find its output and error energy. (iv) Why would the same approach fail for \(w = (0.5, 1)\)?

Answer and worked solution
  1. \(1/(1 + 0.5z) = 1 - 0.5z + 0.25z^2 - 0.125z^3 + \cdots\) → \((1, -0.5, 0.25, -0.125)\).
  2. \((1, -0.5) * (1, 0.5) = (1, 0, -0.25)\). Error energy \(= 0^2 + 0^2 + 0.25^2 = 0.0625\).
  3. Autocorrelation: \(r_0 = 1 + 0.25 = 1.25\), \(r_1 = 0.5\). Desired spike gives \(g = (w_0, 0) = (1, 0)\). Normal equations: \(1.25f_0 + 0.5f_1 = 1\) and \(0.5f_0 + 1.25f_1 = 0\); determinant \(= 1.5625 - 0.25 = 1.3125\). \(f_0 = 1.25/1.3125 = 0.952\), \(f_1 = -0.5/1.3125 = -0.381\). Output \(f * w = (0.952,\ 0.476 - 0.381,\ -0.190) = (0.952, 0.095, -0.190)\). Error energy \(= 0.048^2 + 0.095^2 + 0.190^2 = 0.0023 + 0.0091 + 0.0363 = 0.0476\), smaller than 0.0625: the least-squares filter spreads the error better than simple truncation.
  4. \((0.5, 1)\) is maximum-phase (the larger sample comes second). Its causal inverse \(2(1 - 2z + 4z^2 - \cdots)\) diverges. It has the same autocorrelation as \((1, 0.5)\), so statistical deconvolution, which sees only the autocorrelation and assumes minimum phase, would design the inverse of \((1, 0.5)\) instead and leave a phase error in the output.

Q2 (practice). Marine data show strong ringing. The water is 75 m deep (\(V = 1500\) m/s) and the sea-floor reflection coefficient is 0.5.

(i) What is the reverberation period? (ii) Write the first four coefficients of the reverberation train for a unit spike. (iii) Give the inverse (Backus) filter and verify that it cancels the train up to the fourth term. (iv) What prediction lag would you use in predictive deconvolution?

Answer and worked solution
  1. Water two-way time \(\tau = 2\times75/1500 = 0.100\) s = 100 ms.
  2. \(1 - 2Rz + 3R^2z^2 - 4R^3z^3\) with \(R = 0.5\): \((1, -1.0, 0.75, -0.5)\) at 0, 100, 200, 300 ms.
  3. \((1 + Rz)^2 = 1 + 2Rz + R^2z^2\) → \((1, 1.0, 0.25)\) at 0, 100, 200 ms. Convolution with \((1, -1, 0.75, -0.5)\):
    • lag 0: 1;
    • lag 1: \(-1 + 1 = 0\);
    • lag 2: \(0.75 - 1 + 0.25 = 0\);
    • lag 3: \(-0.5 + 0.75 - 0.25 = 0\). Output is a spike.
  4. A prediction lag equal to the water two-way time, 100 ms, with an operator long enough to span several periods.

Q3 (practice). Consider these statements about migration:

(A) Stolt (f-k) migration is exact for a constant-velocity medium. (B) Using a migration velocity that is too high produces upward-concave “smiles”. (C) Kirchhoff migration sums energy along diffraction hyperbolae and places it at their apices. (D) Post-stack time migration correctly images reflectors beneath a salt body with strong lateral velocity contrast.

Which are correct?

  • (a) A, B and C only
  • (b) A and D only
  • (c) B, C and D only
  • (d) A, B, C and D
Answer and worked solution

(a) A, B and C.

  • (A) Correct; for \(v(z)\) it needs approximations (stretching).
  • (B) Correct: over-migration spreads each point into a smile.
  • (C) Correct: that is diffraction-summation migration.
  • (D) Wrong: strong lateral velocity variation bends rays and moves image points in ways time migration cannot handle; pre-stack depth migration with a good interval-velocity model is needed.

Quick Revision Table

ItemFact
Convolutional model\(x = w * r + n\)
Minimum phase (2-term)\(\lvert w_0\rvert > \lvert w_1\rvert\); causal stable inverse
StaticsTime-invariant; elevation, weathering, residual (surface-consistent)
GainSpherical divergence keeps relative amplitude; AGC does not
CMP vs CDPSame only for flat layers; DMO for dip
Semblance0 to 1; picks increase with time; multiples at lower velocity
NMO stretch\(\Delta f/f \approx \Delta t_{NMO}/t_0\); mute
Stack\(\sqrt N\) for random noise; multiples via residual moveout
Wiener filterToeplitz autocorrelation matrix; Levinson recursion
Spiking deconWhite reflectivity, minimum phase, stationary, low noise
Predictive deconLag \(\alpha\) = multiple period; spiking = lag 1 sample
Backus filter\((1 + Rz)^2\) for water reverberation
F-K\(V_a = f/k\); slow noise near \(k\)-axis; \(k_N = 1/(2\Delta x)\)
Migrator’s equation\(\sin\theta_m = \tan\theta_u\)
Migration shift\(d_x = V^2t(\Delta t/\Delta x)/4\), up-dip
Exploding reflectorOne-way, half velocity
MethodsKirchhoff, FD, Stolt (constant V), phase-shift (\(v(z)\)), RTM
Velocity errorToo high → smiles; too low → frowns

Key Takeaways

What’s Next