|
<< Click to Display Table of Contents >> Navigation: »No topics above this level« Deconvolution |
The Deconvolution module applies Wiener-Levinson predictive deconvolution to seismic gathers. Its primary purpose is to compress the seismic wavelet, remove short-period multiples and reverberations, and broaden the frequency spectrum — all of which improve temporal resolution and make it easier to distinguish individual reflectors in the processed data.
Predictive deconvolution works by computing an autocorrelation of each trace within a user-defined time window (horizon), then solving the Wiener-Hopf equations to design an optimum Wiener filter. This filter predicts the repetitive part of the wavelet (multiples and reverberations) and subtracts it from the trace, leaving behind the underlying reflection series. The key parameters that control the deconvolution are the predictive interval (the lag at which the filter begins to predict) and the operator length (which determines how much of the wavelet's autocorrelation is used).
You can define multiple time zones (horizons) within a single gather, each with its own operator design window and parameters. This allows the deconvolution to adapt to changing wavelet character with depth — for example, using a shorter predictive interval in the shallow section where multiples are closely spaced, and a longer interval in the deeper section. An optional Balance pre-conditioning step can be applied before deconvolution to equalise trace amplitudes and improve operator stability.
For users who need sparser-spike output (wavelet compression beyond what standard Wiener decon achieves), an optional L1-norm iterative solver is available. This mode trades computation time for a more impulsive result, and is suited to cases where the subsurface reflectivity is expected to be sparse.
Connect the seismic gather to be deconvolved. This can be a CMP gather, a shot gather, or any pre-stack or post-stack dataset. The gather must be in the time domain. The module processes one gather at a time and produces an output gather of identical geometry.
Sets the length of the cosine taper (in seconds) applied at the boundaries between adjacent horizon windows. Default: 0.10 s. The taper blends the deconvolution output smoothly across window edges, preventing amplitude discontinuities at the transition between zones. Increase this value if you see sharp amplitude steps at horizon boundaries in the output; reduce it if the transition zone is consuming too much of a thin time window.
When enabled, the autocorrelation used to design the Wiener operator is computed over the entire trace length rather than within the Start Time to End Time window of each horizon. Default: off. Enable this option when the signal-to-noise ratio within the design window is too low to produce a reliable operator estimate, but the full-trace statistic is more representative of the wavelet.
The Horizons collection defines the time zones in which deconvolution operators are designed and applied. You must add at least one horizon. Multiple horizons allow you to vary the deconvolution parameters with depth, adapting to changes in wavelet character and multiple content. Each horizon entry contains the following parameters:
The start time (in seconds) of the operator design and application window for this horizon. Default: 0 s. This value must be less than Time end. Typically set to just above the first arrival or the top of the zone of interest. The autocorrelation of the trace is computed within this window to design the Wiener operator.
The end time (in seconds) of the operator design and application window for this horizon. Default: 5 s. This value must be greater than Time start. The design window should be long enough (typically at least 0.5–1 s) to provide a statistically stable autocorrelation for operator design. The module will report an error if the window length is smaller than the Predictive interval.
The time lag (in seconds) at which the Wiener filter begins to predict the signal. Default: 0.02 s. This is the most important parameter for controlling multiple suppression. A predictive interval equal to the dominant period of the primary wavelet (approximately 1/peak frequency) will compress the wavelet and suppress all multiples with a period longer than the predictive interval. Shorter predictive intervals suppress shorter-period multiples and produce stronger wavelet compression; very short intervals approach spike deconvolution. The predictive interval must be less than the operator Length and less than the window duration (Time end minus Time start).
The prewhitening level, expressed as a fraction of the zero-lag autocorrelation value. Default: 0.05 (5%). Prewhitening adds a small amount of white noise to the autocorrelation before solving the Wiener-Hopf equations, which stabilises the inversion and prevents excessive amplification of frequencies with low signal energy. Higher values (e.g. 10%) produce a more stable but less aggressive operator, useful for noisy data. Lower values (e.g. 1%) produce a sharper, more aggressive operator, suitable for clean, high signal-to-noise data. Values below 0.1% are not recommended as they may produce unstable operators.
The total length (in seconds) of the Wiener prediction error filter. Default: 0.2 s, minimum 0. The operator length must be greater than the Predictive interval. Longer operators capture more of the autocorrelation structure and can suppress longer-period multiples, but also increase computation time. A typical choice is 3–5 times the dominant wavelet period. Excessively long operators risk fitting noise in the autocorrelation rather than the signal.
The number of neighbouring traces to mix (average) when computing the autocorrelation for operator design. Default: 1 (no mixing), minimum 1. Increasing this value averages the autocorrelation over several adjacent traces before solving the Wiener equations, which produces a more spatially consistent operator. This is useful when individual traces are noisy and the per-trace operator estimate is unreliable. Use values of 3–9 for noisy pre-stack data, while keeping a value of 1 for clean post-stack data where trace-by-trace operators are preferred.
The velocity (in m/s) used to compute a hyperbolic mute curve applied to the operator design window of each horizon. Default: 8000 m/s. At this default value, the mute is very mild and excludes very little data from the design window. Reduce this velocity to a value close to the expected stacking velocity in order to mute the refracted first-arrival zone and confine the operator design to the reflection zone only. This prevents the strong first arrivals from dominating the autocorrelation and biasing the operator. The mute curve is displayed as an overlay on the input gather vista.
When enabled, the module computes and outputs the autocorrelation gathers of both the input and output data for quality control purposes. Default: off. Comparing the autocorrelation of the input gather (which will show peaks at the wavelet period and at multiple intervals) with the autocorrelation of the output gather (which should approach a spike at zero lag with no secondary peaks) is the standard method for verifying that deconvolution has succeeded. Enable this option during parameter optimisation and disable it during production runs to save time.
The length (in seconds) of the autocorrelation window used to compute the QC autocorrelation output gathers. Default: 0.5 s, minimum 0. This parameter is only active when Autocorrelation is enabled. Set this to cover at least the expected period of the primary wavelet and the first few multiple reflections. A value of 0.3–0.5 s is suitable for most surface seismic datasets.
When enabled, a trace balance (amplitude equalisation) step is applied to the data before deconvolution. Default: off. The balance normalises each trace amplitude within the specified Start window to End window time range, then designs and applies the deconvolution operator on the normalised data. After deconvolution, the amplitude scaling derived from the balance step is reapplied to restore relative amplitude relationships. This pre-conditioning step helps stabilise the Wiener operator when there are large amplitude variations across traces (e.g., due to source-receiver coupling variations or geometrical spreading), ensuring that the operator design is driven by wavelet character rather than amplitude differences.
The start time (in seconds) of the amplitude window used by the Balance pre-conditioning step. Default: 0.01 s. Only active when Using Balance is enabled. Set this to the beginning of the reflection zone you want to equalise — typically just after the first arrival mute.
The end time (in seconds) of the amplitude window used by the Balance pre-conditioning step. Default: 0.05 s. Only active when Using Balance is enabled. Set this to the end of the time range over which you want to equalise trace amplitudes before designing the deconvolution operator.
When enabled, substitutes the standard Wiener-Hopf least-squares (L2) solver with an iterative L1-norm sparse solver. Default: off. The L1 solver promotes sparsity in the deconvolved output, meaning it seeks a solution consisting of as few strong spikes as possible. This can produce superior wavelet compression on datasets where the reflectivity is genuinely sparse (e.g., hard-kick carbonate sequences or evaporite sections). It is significantly more computationally expensive than the standard L2 solver and is not recommended for routine production processing. When this option is enabled, the Number of L1 solver iterations and Sparsity damping parameters become available.
The maximum number of iterations the L1 sparse solver will perform per trace. Default: 100, minimum 1. Only active when Use L1 solver is enabled. More iterations allow the solver to converge to a sparser solution but increase computation time proportionally. For most datasets, 50–200 iterations is sufficient. Increase this value if the output does not appear sparse enough; reduce it to speed up processing at the cost of solution quality.
A small regularisation constant that prevents the L1 solver from becoming numerically unstable in the presence of very small residuals. Default: 0.0001, minimum 1e-9. Only active when Use L1 solver is enabled. This parameter controls the balance between sparsity and stability: smaller values push the solution toward greater sparsity but risk numerical instability, while larger values stabilise the solver at the cost of a less sparse result. The default value is appropriate for most cases; adjust only if convergence problems are observed.
The deconvolved output gather. It has the same geometry (number of traces, sample interval, and record length) as the input gather. Short-period multiples and reverberations present in the input should be attenuated, and the seismic wavelet should be compressed toward a broader-band, more impulsive waveform. If the Calculate difference option is enabled (inherited from the standard sequence procedure), a difference gather is also produced showing what was removed.
A QC gather showing the autocorrelation of the input traces, computed using the Time length window. Only populated when Autocorrelation is enabled. Use this gather to identify the primary wavelet period (the first side-lobe of the zero-lag peak) and the period of dominant multiples (secondary peaks at longer lags). This information guides the selection of the Predictive interval and operator Length parameters.
A QC gather showing the autocorrelation of the deconvolved output traces, computed using the Time length window. Only populated when Autocorrelation is enabled. A well-deconvolved result will show a near-perfect spike at zero lag with minimal energy at all other lags, confirming that the wavelet has been compressed and that multiple energy has been suppressed. Residual peaks at non-zero lags indicate that the deconvolution is incomplete, and the parameters may need adjustment.