Kurtosis

Key Takeaways

• Kurtosis as a seismic trace attribute is computed from the amplitude distribution of a seismic trace window (typically a 100 to 500 millisecond window centered on the zone of interest) and used to characterize the statistical character of the reflectivity series convolved into the trace: a trace dominated by a single large-amplitude reflector (such as a bright spot from a gas-bearing reservoir against background shale) will have a highly leptokurtic amplitude distribution (high kurtosis) because the amplitude distribution is dominated by a few large values at the tails and many near-zero values in the center, resembling a sparse impulse series; a trace dominated by many small-amplitude interfering reflectors from interbedded lithologies (the background clastic section) will have a nearly Gaussian amplitude distribution (kurtosis near 3) or slightly leptokurtic; kurtosis attributes extracted from seismic data can be used to map the spatial distribution of sparse reflectivity (which often correlates with gas-bearing sands, carbonates, or other lithological anomalies) and to guide the design of sparse-spike deconvolution algorithms that assume a non-Gaussian (high-kurtosis) reflectivity series, as the assumption of a leptokurtic reflectivity distribution enables the separation of the wavelet and reflectivity series by independent component analysis methods that fail when Gaussian statistics are assumed.
• In permeability statistics for reservoir characterization, kurtosis reflects the degree to which the permeability distribution is dominated by extreme high-permeability values (streaks, fractures, dissolution voids, or highly sorted grains) that control flow in a disproportionate fraction of the pore volume: log-normal permeability distributions (which are typical of many fluvial, deltaic, and turbidite sandstone reservoirs) have excess kurtosis ranging from 0 to 3 (nearly Gaussian on the log scale), while fracture-dominated or vug-dominated carbonate reservoirs frequently exhibit highly leptokurtic permeability distributions (excess kurtosis of 10 to 100) because a small number of fractures or vugs have permeabilities thousands of times higher than the matrix; the practical engineering consequence of high permeability kurtosis is that fluid injection (in waterfloods and EOR programs) preferentially channels through the high-permeability tails of the distribution, sweeping a small fraction of the pore volume at very high velocity while the bulk of the pore volume at lower permeabilities is bypassed, reducing displacement efficiency and increasing water production at producing wells far sooner than simple average permeability would predict; permeability kurtosis estimated from core plug measurements can be compared to kurtosis estimated from well test data to assess the degree to which small-scale permeability heterogeneity (captured by core plugs) scales up to the well-scale flow behavior (characterized by pressure transient tests).
• Independent Component Analysis (ICA) uses kurtosis as the objective function to separate mixed seismic or well log signals into statistically independent source components: ICA algorithms find linear combinations of observed mixed signals that maximize the non-Gaussianity (measured by kurtosis or a related quantity such as negentropy) of the extracted components, based on the central limit theorem principle that the sum of independent random variables tends toward Gaussian (low kurtosis), so the most non-Gaussian (high kurtosis) linear combinations are the best estimates of the original independent source signals; in seismic processing, ICA-based blind deconvolution uses the high-kurtosis assumption for the reflectivity series to separate the seismic wavelet (approximately Gaussian due to the band-limited filtering of the Earth) from the reflectivity series (impulsive and non-Gaussian) without knowing the wavelet in advance; in multivariate log analysis, ICA separates composite well log signatures into lithological and fluid-related components that are more geologically interpretable than the principal components of standard PCA.
• Moment calculation and statistical inference challenges arise when applying kurtosis to small petroleum data samples: the kurtosis estimator (the sample fourth central moment divided by the square of the sample variance) is highly sensitive to outlier values because extreme observations contribute to the fourth power in the numerator, making sample kurtosis very imprecise for small datasets; for a sample of 30 observations, the standard error of the kurtosis estimator is approximately 1.4 (meaning the kurtosis would need to exceed 2.8 to be statistically distinguishable from normal at 95 percent confidence), so kurtosis analysis of small well core datasets (which may have only 20 to 50 plugs per reservoir interval) can be unreliable; robust alternatives to standard kurtosis include the L-kurtosis (a linear combination of order statistics that is less sensitive to extreme values) and the quantile-based kurtosis (computed from the ratio of inter-quartile ranges), both of which can be applied to small samples with more reliable statistical properties; for seismic trace attribute kurtosis, the large number of samples in the trace window (hundreds to thousands of samples at typical sampling rates) makes the kurtosis estimator much more reliable than for core data.
• Production data kurtosis in field development analysis provides insight into the distribution of well productivity: the kurtosis of the initial production rate (IP) distribution across all wells in a shale play or conventional field characterizes how skewed and tail-heavy the productivity distribution is; high kurtosis in IP distributions (common in unconventional plays where a few high-IP wells dominate the EUR distribution) indicates that development economics are sensitive to the proportion of high-IP wells drilled, and that average IP statistics are poor predictors of individual well economics because the distribution is far from symmetric; Monte Carlo reserve estimation models calibrated to high-kurtosis IP distributions will produce reserve estimates with heavier tails (higher P90/P10 ratios) than models calibrated to normal distributions with the same mean, correctly reflecting the geological uncertainty in which wells will intercept the highest-permeability sweet spots in the play; production decline kurtosis analysis can also detect non-hyperbolic decline behavior (exponential or harmonic) by comparing the kurtosis of the production rate distribution at different time steps to the expected evolution under each decline model.