Phase variance as a seismic quality-control attribute

This paper introduces "phase variance," a data-driven, frequency-dependent quality-control attribute based on circular statistics that quantifies localized phase dispersion in seismic data without requiring phase unwrapping or global wavelet models, thereby enabling automated assessment of phase reliability to support advanced workflows like AVO and full-waveform inversion.

The Gist

Here is an explanation of the paper using simple language and creative analogies.

The Big Problem: The "Static" on Your Radio

Imagine you are trying to listen to a radio station, but the signal is passing through a thick, bumpy forest before it reaches your antenna. The trees (the Earth's near-surface rocks) scatter the sound waves. Sometimes they delay them, sometimes they speed them up, and sometimes they twist the sound in weird ways.

In the world of oil and gas exploration, geophysicists use sound waves (seismic data) to "see" underground. But just like your radio, the land-based signals get distorted by the messy ground above the oil or gas.

The Old Way of Fixing It:
For decades, geophysicists have tried to fix this using a method called "Surface-Consistent Deconvolution." Think of this like trying to fix a blurry photo by applying a single, generic filter to the whole picture. They assume the distortion is the same for every shot and every receiver.

• The Flaw: This works okay for big, slow waves, but it fails miserably at high frequencies (sharp details). It's like trying to fix a specific scratch on a lens by smearing the whole image. It also doesn't tell you how bad the blur actually is; you just have to squint at the screen and guess, "Hmm, that looks a little clearer, maybe?"

The New Idea: The "Crowd Chorus"

The authors of this paper propose a new way to look at the problem. Instead of looking at one single sound wave at a time, they look at a group (an ensemble) of waves together.

Imagine a choir.

• Perfect Signal: If the choir is singing in perfect harmony, every voice is on the exact same note and timing.
• Noise: If the choir is just a crowd of people talking randomly, there is no single note; it's just chaos.
• The Real World: In the middle, some people are singing the right note, but others are slightly off-key or late.

The paper introduces a new tool called Phase Variance. Instead of asking "Is the volume loud enough?" (which is what old methods check), it asks, "How much are the voices in the choir disagreeing with each other?"

How It Works: The "Circle" Analogy

Seismic waves are cyclical (like a clock face). A phase of 0 degrees is the same as 360 degrees.

• The Old Mistake: If you have a clock showing 11:59 and another showing 12:01, a normal math calculator might think they are 2 hours apart because it treats them as straight lines. But on a circle, they are right next to each other!
• The New Solution: The authors use Circular Statistics. They treat the data like points on a dartboard or a compass.
  • If all the darts are clustered tightly in one spot, the "Phase Variance" is low. This means the signal is strong and reliable.
  • If the darts are scattered all over the board, the "Phase Variance" is high. This means the signal is messy and unreliable.

The "Aha!" Moment: Loud Doesn't Mean Good

The most surprising discovery in the paper is about Volume vs. Clarity.

Imagine a speaker playing music.

• Old Method: They turn up the volume (amplitude) to make the music sound better. They see the waves getting bigger and think, "Great! We found more high-frequency details!"
• New Method: The authors turn up the volume, but they also check the "Phase Variance." They find that while the volume is loud, the "voices" in the choir are still screaming at different pitches. The signal is loud, but it's garbage.

The Analogy: It's like a crowded party where everyone is shouting (high amplitude). If you turn up the microphone, it sounds even louder, but you still can't understand what anyone is saying because they are all talking over each other (high phase variance). The old methods thought the party was getting clearer just because it got louder. The new method realizes, "No, it's just a louder mess."

Why This Matters

1. Honest Quality Control: This new tool gives a number (0 to 1) that tells you exactly how reliable the data is at every specific frequency.
   • 0 = Perfect harmony (Trust this data!).
   • 1 = Total chaos (Ignore this data!).
2. Saving Money and Time: In the past, geophysicists might try to use high-frequency data to find small oil pockets, only to realize later that the data was too noisy to be useful. This tool tells them before they start the expensive work: "Hey, the data is only reliable up to 25 Hz. Don't waste time trying to use the 50 Hz data."
3. Better Imaging: By knowing exactly where the "noise" starts, they can use special filters to clean up the signal, leading to much sharper images of the underground.

Summary

This paper is about changing how we measure the quality of seismic data.

• Old Way: "Is it loud? Does it look okay to my eye?" (Subjective and often wrong).
• New Way: "Are the waves in the group agreeing with each other?" (Objective and mathematical).

It's like switching from judging a choir by how loud they are, to judging them by how well they are singing in tune. This helps geologists stop wasting time on "loud but messy" data and focus only on the "clear and harmonious" signals that actually reveal what's underground.

Technical Summary

Here is a detailed technical summary of the paper "Phase variance as a seismic quality-control attribute" by Akshika Rohatgi, Andrey Bakulin, and Sergey Fomel.

1. Problem Statement

Land seismic data recorded on the surface is heavily distorted by near-surface heterogeneity (small-scale elastic contrasts). These distortions introduce trace-specific, frequency-dependent phase perturbations that persist even after advanced time processing (e.g., statics and deconvolution).

• Limitations of Conventional Processing: Standard workflows rely on surface-consistent deconvolution, which assumes that near-surface distortions are repeatable across traces for a given source or receiver. This approximation fails to correct localized, non-surface-consistent phase distortions common in modern point-receiver (single-sensor) data.
• Lack of Quantitative Metrics: There is currently no direct, quantitative measure of phase reliability. Quality control (QC) relies on indirect proxies like amplitude behavior or visual inspection, leaving residual phase disorder undiagnosed.
• The "Speckle" Regime: In highly scattering environments (e.g., deserts, basalts), forward scattering creates a "speckle-like" regime where individual traces appear chaotic, yet ensemble statistics remain stable. Conventional methods cannot effectively characterize or correct this frequency-dependent phase randomization.

2. Methodology

The authors propose a stochastic framework that treats seismic phase not as a linear trace attribute, but as a circular random variable. They apply circular statistics to local ensembles of seismic traces to quantify phase dispersion.

Core Concepts

• Circular Statistics: Since seismic phase is wrapped in the interval $[-\pi, \pi]$, linear averaging is invalid (e.g., averaging $179^\circ$and$-179^\circ$yields$0^\circ$, which is incorrect). The method uses unit phasors to represent phase.
• Circular Mean ($\bar{\theta}$): The direction of the resultant vector formed by summing normalized unit phasors. It represents the dominant phase direction of the ensemble.
• Mean Resultant Length ($\bar{R}$): A measure of phase alignment (clustering strength), ranging from 0 (uniform/random) to 1 (perfectly aligned).
• Phase Variance ($V$): The primary attribute introduced, defined as:
  $$V(\omega) = 1 - \bar{R}(\omega)$$
  • $V = 0$: Perfect phase coherence (all traces aligned).
  • $V = 1$: Complete phase incoherence (uniform distribution, noise-dominated).

Workflow

1. Ensemble Selection: Select a local window of $N$ traces (e.g., 2,000 traces) within a prestack gather.
2. Frequency Domain Transformation: Convert traces to the frequency domain ($X_k(\omega)$).
3. Normalization: Normalize spectra to unit magnitude to isolate phase information ($e^{i\theta_k(\omega)}$), removing amplitude bias.
4. Calculation: Compute the circular mean and resultant length for each frequency $\omega$ within the window.
5. Mapping: Assign the calculated $V(\omega)$ to the window center, generating a phase-variance map across time, offset, and frequency.

The method operates without phase unwrapping, making it robust in noisy or incoherent regions where unwrapping algorithms typically fail.

3. Key Contributions

• Introduction of Phase Variance ($V$): A new, automated, frequency-by-frequency QC attribute that quantifies localized phase dispersion.
• Stochastic Framework: Reframing seismic phase analysis from single-trace estimation to ensemble-based circular statistics, providing a physically grounded measure of scattering-induced noise.
• Decoupling Phase from Amplitude: The method isolates phase behavior, revealing that amplitude-based metrics (like SNR or bandwidth extension) can be misleading if phase coherence is not preserved.
• Frequency-Dependent Analysis: The ability to identify specific frequency bands where phase is reliable versus those dominated by noise, which is critical for phase-sensitive workflows.

4. Results

The paper validates the method through synthetic tests and field data application:

Synthetic Tests

• Controlled experiments with imposed phase perturbations (modeled via von Mises distributions) confirmed that the calculated circular variance accurately recovers the known level of phase disorder.
• The metric successfully distinguished between additive noise (broadening distributions) and multiplicative scattering noise (frequency-dependent broadening).

Field Data Application (Land Prestack Gathers)

• Processing Impact: Conventional time processing (statics, deconvolution, noise attenuation) significantly reduced phase variance in the low-to-intermediate frequency range and at mid-to-far offsets.
• Limitations Revealed:
  • High Frequencies: Despite apparent amplitude enhancement and bandwidth extension in processed data, phase variance remained high ($V \approx 1$) at high frequencies, indicating the data is still noise-dominated.
  • Near Offsets (Noise Cone): Phase variance remained elevated in near-offset traces regardless of processing, highlighting the limits of surface-consistent corrections in these zones.
• Comparison with SNR: While Stack-based Signal-to-Noise Ratio (SNR) showed overall improvement, it masked the fact that high-frequency phase coherence was not recovered. Phase variance provided a more honest assessment of usable bandwidth.

5. Significance and Implications

• Redefining Effective Bandwidth: The authors argue that "usable bandwidth" should be defined by phase coherence (low $V$) rather than just spectral amplitude. This prevents the use of high-frequency data that appears strong in amplitude but is phase-randomized.
• Guiding Processing Workflows: Phase variance serves as an objective metric to:
  • Evaluate the efficacy of specific processing steps.
  • Guide parameter selection for noise attenuation and deconvolution.
  • Identify where phase distortions persist and require specialized correction (e.g., time-frequency masking).
• Enabling Advanced Applications: Reliable phase data is a prerequisite for AVO analysis, migration, and Full-Waveform Inversion (FWI). By identifying the frequency/offset ranges where $V$ is low, practitioners can confidently apply these phase-sensitive techniques.
• Future-Proofing: As seismic acquisition moves toward denser sampling and higher frequencies, the ability to statistically characterize and manage phase variability becomes increasingly critical. This framework provides the statistical foundation for next-generation phase-conditioning algorithms.

In summary, the paper establishes Phase Variance as a fundamental, model-free diagnostic tool that shifts seismic QC from visual/amplitude-based heuristics to a rigorous, statistical assessment of phase reliability.