Semblance Analysis¶
A moveout-velocity scan that scores cross-trace coherence after trial alignment to identify plausible propagation-velocity trends.
Core Idea¶
Semblance analysis selects plausible propagation velocities by asking which trial moveout correction makes an event line up across traces recorded at different source–receiver offsets. For each candidate velocity, a common-midpoint gather is corrected, agreement among its traces is measured in a short time window, and the score is plotted by time and trial velocity. A coherent ridge or peak guides a stacking-velocity pick, which can then be checked against surrounding data and used in later correction and stacking. The method is a comparison across candidate models, not simply a filter applied once to a trace.[1][2]
In conventional semblance, the score is the sum over a time window of the squared stack of the corrected traces, divided by the number of traces times the sum of their individual squared amplitudes. For a nonzero denominator it lies between zero and one: perfect agreement can attain one, while cancellation reduces the numerator. It is a measure of alignment under a moveout model, not a certificate that a reflection is primary or that the chosen velocity equals a layer's interval velocity.[2][3]
Structural Signature¶
Sig role-phrases:
- Multi-offset gather — Several traces sampling a candidate common event at different offsets supply the comparison. Without redundancy, the cross-trace test has no useful meaning.[1]
- Trial moveout function — A candidate velocity predicts how arrival time varies with offset. Conventional CMP processing often uses an approximately hyperbolic normal-moveout model; the assumed geometry must be stated.[1]
- Alignment operation — The trial correction attempts to flatten the event before traces are compared. A velocity that leaves an event curved generally produces less coherent stacking.[1]
- Local normalized score — A short-window ratio of squared stack energy to individual-trace energy assesses whether corrected amplitudes agree. Zero-energy windows require explicit handling rather than being assigned a meaningful zero-to-one score.[2]
- Time–velocity spectrum and pick — Scores across trials form a velocity spectrum; a geologically and spatially plausible trend is chosen, not an isolated high pixel taken as truth.[1][3]
- Model-validity checks — Amplitude variation, polarity, noise, nonhyperbolic moveout and survey geometry may shift or suppress peaks; variants change the score or moveout family.[4]
Condensed: multi-offset observations + trial alignment + normalized trace agreement + cross-trial interpretation → a provisional moveout-velocity trend.
What It Is Not¶
- Not a direct measurement of every layer's interval velocity. A picked moveout/stacking velocity can approximate an RMS velocity under suitable assumptions. Deriving interval velocities requires an additional model or inversion such as Dix's equation, which amplifies errors in the picks.[3]
- Not automatic removal of all noise or multiples. Noise may lower or distort peaks, and a coherent unwanted event can itself score highly. The score guides interpretation and later processing; it does not prove event origin.[2]
- Not generic normal-moveout correction. Correction uses a selected velocity; semblance analysis compares many trial corrections to help select it.[1]
- Not all seismic inversion or structural seismic analysis. The former is a broad class of model-estimation problems; the live Seismic Analysis entry concerns earthquake response of engineered structures, not reflection-data velocity scanning.
- Not any coherence calculation. A coherence image without time–velocity trial scanning omits this method's decision structure.
Scope of Application¶
In reflection seismology, the input is commonly a CMP gather. A suite of candidate velocities is used for NMO correction, and semblance is computed over corrected traces. The resulting time–velocity panel helps choose a trend for stacking or further imaging. The conventional hyperbolic approximation is useful only to the extent that it matches the observed moveout; lateral heterogeneity, anisotropy, long offsets or complex wavepaths may call for a richer model.[1][2]
Taner and Koehler's 1969 velocity-spectra paper anchors the historical method; the exact normalized conventional semblance expression used here is checked against a later author-hosted method, not inferred from the inaccessible full original paper.[5][2]
The same logic has been adapted to ground-penetrating-radar CMP data, where the sought velocity is electromagnetic rather than seismic. Jacob and Urban compared hand picking, cross-correlation and semblance in one archaeological survey: their CMP estimate improved on a problematic common-offset hyperbola fit, but semblance was faster and less precise than the alternatives in that particular comparison. This supports transfer of the method, not universal superiority.[6]
Clarity¶
Imagine several traces in which a reflected pulse arrives later at larger offsets. A trial velocity that predicts too much correction bends the corrected events one way; one that predicts too little leaves them bent the other way. At a better trial velocity, their peaks line up in the chosen window, their amplitudes add before squaring, and semblance rises. Scan many trials and repeat across zero-offset times to form the panel.[1][2]
The normalized ratio matters: a large unnormalized stack can reflect large input amplitudes rather than unusually good agreement. Conversely, two real events with opposite polarities can cancel in a simple stack and receive a low conventional score. Fomel's AB semblance was developed for prestack gathers with amplitude trends; changing the score can preserve a different kind of coherence but also changes its sensitivity and resolution.[2][4]
Manages Complexity¶
The analysis reduces a gather of many offset-dependent waveforms to a two-dimensional time–velocity decision surface. Instead of manually following every curve on every trace, an interpreter sees where one moveout model aligns redundant measurements. The reduction is operationally useful because subsequent stacking can reinforce aligned signal relative to uncorrelated contributions.[1]
This compression has a cost. Different geologic events, coherent noise, limited fold, amplitude trends and imperfect correction can compete in the same spectrum. A picked ridge should be evaluated with neighboring gathers, model assumptions and independent constraints; the method does not make the inverse problem unique.[4][3]
Abstract Reasoning¶
Specify a gather, time window and candidate moveout family. For each candidate velocity, map every trace to the time at which that model predicts the same event, sample or correct the amplitudes, then calculate the normalized agreement. Compare scores across velocities and times. A peak is evidence that this model organizes the observed arrivals, subject to data quality and model fit; it is not itself evidence that the same numerical value holds within one physical layer.[2][1]
If the numerator is high while individual-trace energy is low, the ratio can still be meaningful only if the denominator is nonzero and the data are reliable. If the picked trend changes abruptly from gather to gather, ask whether the geology warrants it or whether noise and ambiguous peaks dominate. For an interval-velocity claim, perform and scrutinize the additional inversion rather than relabeling a stacking-velocity pick.[3]
Knowledge Transfer¶
The portable structure is hypothesize alignment, normalize agreement, compare candidate parameters. It carries from acoustic seismic surveys to electromagnetic GPR CMP surveys because both can record offset-dependent arrivals from subsurface reflectors. What does not transfer automatically is the propagation speed range, wave physics, survey geometry, depth conversion or precision of a particular picking procedure.[6][1]
This is a domain-specific method rather than a prime because the named procedure depends on a multi-offset propagation model and a particular coherence statistic. Its broader skeleton resembles model-guided alignment, but that resemblance does not make every correlation score a semblance velocity analysis.
Examples¶
Constructed three-trace velocity comparison¶
Take a constructed three-trace CMP example with two time samples in the scoring window; these are illustrative corrected amplitudes, not measured field data. At trial velocity \(v_A\), the three NMO-corrected trace vectors are \((1,0),(1,0),(1,0)\). Equation (5) in Gan and colleagues gives numerator \((1+1+1)^2+(0+0+0)^2=9\) and denominator \(N\sum_{j,k}s_{jk}^2=3(3)=9\), hence semblance \(S(v_A)=1\). At competing \(v_B\), suppose the corrected vectors are \((1,0),(0,1),(0,1)\): numerator \(1^2+2^2=5\), denominator again \(3(3)=9\), so \(S(v_B)=5/9\). On this local window the scan selects \(v_A\) over \(v_B\); it has not inferred an interval velocity or proven that the aligned event is a primary reflection. Actual moveout timing, window choice and neighboring gathers would still govern a real pick.[2][1]
Mapped back: gather = three stipulated corrected traces; trials = \(v_A\) and \(v_B\); alignment = all peaks at first sample for \(v_A\) versus split peaks for \(v_B\); score = \(9/9\) versus \(5/9\); output = local preference for \(v_A\), not a field velocity measurement.
Archaeological ground-penetrating radar¶
Jacob and Urban studied a small iron-working site in Rhode Island. Their common-offset hyperbola fit gave unexpectedly high velocities, while a CMP analysis gave a more reasonable estimate. Within the CMP comparison, semblance was faster than hand-picking and cross-correlation but less precise. Thus a concrete method decision in that reported comparison is to treat the quick semblance scan as a provisional velocity guide and use a more precise CMP procedure when the depth estimate requires it. The accessible abstract reports these relative outcomes, not the underlying gather amplitudes or numerical velocities; none are invented here.[6]
Mapped back: gather = reported CMP radar data at a Rhode Island iron-working site; trial = radar-wave velocity candidates; output = more reasonable CMP estimate than the common-offset hyperbola fit; consequence = faster semblance but lower precision than the other CMP procedures; limit = abstract-only relative findings, no numerical reconstruction.
Near miss: single-trace frequency filtering¶
A band-pass filter may make one trace easier to view, but it neither predicts relative arrivals across offsets nor scans trial velocities. It may precede a semblance analysis; it is not itself one.
Structural Tensions¶
The difference between a coherence peak and geologic truth is an interpretation boundary, not an inherent opposed-cost tension: even \(S=1\) in the constructed example certifies only equality of sampled amplitudes after a specified correction. Diagnostic: does the picked trend persist across neighboring gathers and independent constraints, and is the moveout model plausible?[1][3]
Simple constant-amplitude agreement versus real amplitude behavior. Conventional semblance rewards a common waveform across traces, but amplitude variation with offset or polarity change can lower its score for a real event. Trend-aware variants may help at the cost of a changed statistic and sensitivity. Diagnostic: is a weak score evidence of poor alignment or a mismatch between the score's amplitude model and the data?[4]
Structural–Framed Character¶
Semblance analysis is a structural signal-processing operation framed by a propagation model. Its ratio is mathematical and not itself an evaluative judgment about a geological interpretation; the value judgment enters when a user calls a peak useful or trustworthy. Human practice selects trace gathers, trial moveout families, windows and plausible events, while survey geometry and acquisition conditions constrain those choices. The method arose in seismic velocity-analysis practice and has a reported radar application; no institution's label makes an arbitrary coherence score a velocity estimate.[1][6] The same formula can highlight a coherent nuisance event, so the human or automated pick remains provisional.[2]
Words such as alignment and coherence travel to other signals, but their appearance alone does not identify this velocity-analysis method. Recognition requires trial propagation moveout and a scored cross-record response; importing the ratio without the model is an analogy or a different measurement. Its character: a model-guided scan whose score is exact for a chosen window but whose velocity interpretation is conditional on physical and survey assumptions.
Structural Core vs. Domain Accent¶
The structural skeleton is repeated parameter testing by alignment and normalized cross-record agreement. The actual Measurement Method parent carries the broader controlled-observation-to-inference chain, but does not supply this particular alignment/coherence skeleton. Whether that pattern forms a separate portable prime is a future-prime question, not an asserted parent. The domain-bound mechanism is CMP geometry, travel-time moveout, seismic or radar propagation velocity, and downstream NMO/stacking or depth conversion. Strip away trial moveout and the velocity-spectrum step and one has merely a general coherence measure; strip away source and survey physics and a peak no longer has a justified velocity interpretation. The named method therefore fails the prime bar despite its reusable analysis motif.[1][6]
Instantiates / Related Primes¶
This entry is a kind of Measurement Method.
Measurement Method is the accepted strict genus: semblance analysis aligns trials and scores cross-trace coherence to identify plausible propagation-velocity trends. Similarity and Alignment remain conceptual relatives. A score peak is an inferential candidate, not automatic proof of a primary reflection or true interval velocity.
Relationships to Other Abstractions¶
Current abstraction Semblance Analysis Domain-specific
Parents (1) — more general patterns this builds on
-
Semblance Analysis is a kind of Measurement Method Domain-specific
Trial moveout alignment and normalized cross-trace coherence constitute a specified measurement method, with propagation-velocity scanning as its differentia.The child method maps seismic trace alignment under trial moveout parameters to a measured coherence score, a specialized Measurement Method. Generic measurement methods need not use multitrace semblance or velocity scanning. The peak is an inferential candidate under the method's assumptions, not automatically the physical layer velocity.
Hierarchy path (1) — routes to 1 parentless root
- Semblance Analysis → Measurement Method → Measurement
Neighborhood in Abstraction Space¶
Semblance Analysis sits in a sparse region of the domain-specific corpus (93rd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.
Family — Unclustered & Miscellaneous (2551 abstractions)
Nearest neighbors
- Stabilized Inverse Q Filtering — 0.79
- Reflection Seismology — 0.79
- Infrasonic passive differential spectroscopy — 0.78
- Program Profiling — 0.78
- Pursuit Curve — 0.78
Computed from structural-signature embeddings · 2026-10-08
Not to Be Confused With¶
An NMO velocity is a parameter that flattens an event under a moveout approximation. An interval velocity describes propagation within a layer and is obtained only with additional assumptions or inversion. Semblance is the local normalized agreement statistic; semblance analysis is the repeated use of that statistic while scanning trial moveout models. AB or weighted semblance are variants, not evidence that the conventional measure already handles every amplitude trend.[2][3][4]
References¶
[1] Stanford Exploration Project, Conventional velocity analysis, trial velocity and moveout-correction workflow. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o
[2] Gan and colleagues, author-hosted method for conventional and weighted semblance, equation (5) and surrounding discussion. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l
[3] Stanford Exploration Project, Dix inversion of interval velocity estimation, distinction between RMS picks and interval velocity plus sensitivity to errors. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g
[4] Fomel, original AB semblance study (2009), abstract and theory navigation. registry ↩a ↩b ↩c ↩d ↩e
[5] Taner and Koehler, original velocity-spectra paper (1969). Bibliographic record and abstract checked; full text not checked in this pass. registry ↩
[6] Jacob and Urban, original GPR CMP study (2016), abstract and reported method comparison; full text not checked. registry ↩a ↩b ↩c ↩d ↩e