From shot to stack: processing at a glance

What raw data looks like

A seismic survey fires a source (a vibrator truck, a small explosion, or an air gun) and records the resulting ground motion at an array of receivers (geophones on land, hydrophones in streamers towed behind a ship). Each source-receiver pair produces one trace — a time-series of amplitudes. A single firing of the source gives you one shot record: typically hundreds or thousands of traces.

Raw shot records are dominated by things you don’t want: surface waves (ground roll), noise from the source itself, reflections from nearby boundaries that wrap into your window of interest (multiples), and random noise from wind, waves, traffic, or pumping equipment. The job of processing is to dig the faint reflection signal out of that mess.

CMP and fold

The common midpoint (CMP) of a trace is the surface point exactly halfway between its source and its receiver. Multiple traces can share the same CMP — for example, source at 100 m with receiver at 300 m shares a CMP with source at 150 m and receiver at 250 m. Both traces reflect off the same subsurface point for horizontal layers, just at different take-off angles. The number of traces sharing a given CMP is called the fold. A "60-fold survey" has 60 traces per CMP on average.

When we correct each trace for its different travel path and sum them together, the reflection signal reinforces (it arrives at roughly the same time on each trace after correction) while the noise partially cancels (it is uncorrelated between traces). That operation — correct for moveout, then sum the CMP — is the stack, the single most important noise-reducing operation in the pipeline. Stacking improves signal-to-noise ratio approximately by $\sqrt{N}$ where N is the fold. At 60 fold, you improve SNR by a factor of roughly 8.

Normal moveout and velocity analysis

Within one CMP, the farther the source-receiver offset, the longer the travel path to the reflector and back, so the reflection arrives later. The extra time delay is called normal moveout (NMO). For a horizontal reflector at two-way time $t_0$ under a medium with RMS velocity $V$, the travel time at offset x is:

Correcting each trace so the moveout disappears — flattening the reflection across offset — is the NMO correction. To do it, you need to know the velocity V at every time. The process of determining V is called velocity analysis, and it is typically done by finding the velocity that best flattens coherent events in a CMP gather. An interpreter’s velocity model is the single most consequential input to everything that follows: migration position, time-to-depth conversion, synthetic tie, everything.

When NMO correction uses a velocity that is slightly wrong, reflectors are not perfectly flat after correction — they sag (velocity too high) or arc upward (velocity too low). The resulting stack still works, but the image is softer than it could be. Modern processing iterates the velocity estimate multiple times to tighten the result.

The main processing steps

• Static corrections (land only): remove time shifts caused by variable shallow-surface conditions (differences in near-surface weathering thickness, elevation, etc.). Without these, CMPs don’t align and stacking fails.
• Denoising: specific filters for ground roll, swell noise, ambient noise, coherent linear noise, and others. Each attack one class of unwanted energy.
• Deconvolution: sharpens the wavelet, attempting to reduce a broad source pulse toward a narrower, more impulsive shape. Covered briefly in §1.2; we’ll return to the idea in Part 6.
• Multiple attenuation: removes energy that has reflected more than once (e.g., sea surface + sea floor + sea surface again). Multiples can look convincingly like primary reflections; missing one is a classic interpreter trap. Techniques include radon filtering, wave-equation demultiple, and the surface-related multiple elimination family of methods.
• Migration: repositions energy to its correct spatial location — the topic of §1.4. Can be done before stack (pre-stack) or after (post-stack); pre-stack is more expensive but gives a better image when velocities are complex.
• Stacking: the NMO-and-sum operation we described. The output is one trace per CMP.
• Post-stack enhancements: residual filtering, AGC (automatic gain control), structure-oriented smoothing. Mostly cosmetic — they make the data look cleaner without adding new information.

Signal-to-noise versus bandwidth is always a tradeoff

Almost every processing step you can name trades one of these against the other. Strong filtering improves SNR but kills high frequencies, degrading resolution. Aggressive deconvolution sharpens the wavelet (wider bandwidth) but amplifies noise. Stacking improves SNR but can smear signal if velocities are wrong. There is no "correct" answer; there is only "the right tradeoff for this dataset and this target." Interpreters who understand this can talk to processors about what they actually need.

Pre-stack or post-stack?

Historically, interpreters worked almost exclusively with post-stack data — the single trace per CMP that comes out of stacking. Today, pre-stack analysis (looking at how amplitudes change with offset, for AVO work) is routine for reservoir targets. Pre-stack data is ~60× larger (one trace per offset, per CMP) and needs careful velocity control, but it contains information that stacking throws away. We’ll return to this split explicitly in Part 5 when we study AVO.

For now the takeaway: the post-stack image you first load into your interpretation software is a compressed summary of pre-stack data. Most structural and stratigraphic work runs on the post-stack image. Reservoir and DHI work often needs the pre-stack gathers.