Autotrack: Seismic Horizon Picking, 3D Interpretation, and Algorithms

Key Takeaways

• Tracking criteria and algorithm types: Commercial seismic interpretation packages offer several distinct autotracking algorithms that differ in the criterion used to propagate the pick from trace to trace. Cross-correlation tracking computes the normalized cross-correlation coefficient between the seed waveform and the candidate trace within a defined time window, picking the time shift at maximum correlation; it is robust on continuous, high-amplitude events but sensitive to lateral waveform changes from tuning and interference. Amplitude tracking picks the local maximum or minimum amplitude nearest to a projected time, without requiring waveform shape matching; it is fast but prone to jumping between adjacent peaks when events are closely spaced near tuning thickness. Phase tracking follows a constant phase value (typically zero-crossing or peak phase) through the data, which is useful when amplitude varies laterally but the reflection phase is stable, as in some carbonate pinnacle reef edges. Voxel-based or 3D tracking searches simultaneously in all directions from a seed point through the 3D data volume, visiting each voxel once and accepting it if its value matches the target criteria within tolerance; it is the fastest method on large volumes but may produce diffuse, noisy horizon maps near faults and pinchouts where the algorithm struggles to determine which voxels belong to the target horizon. Machine-learning-based autotracking, available in newer platforms, trains a classifier on interpreter-confirmed picks to predict the horizon probability at unvisited locations, producing more geologically reasonable picks in noisy areas by incorporating regional trend information.
• Seed point strategy and fault handling: The placement and number of seed picks are the most important user decisions in autotracking. A single well-placed seed pick in a clearly identified, high-quality part of the 3D volume is sufficient to initiate tracking on a laterally continuous event, but the resulting pick will halt at any discontinuity (fault, unconformity, noise burst) where waveform similarity drops below the tolerance threshold. To track across a faulted surface, the interpreter must define fault polygons that stop the tracking algorithm at the fault plane, preventing it from jumping across the fault to the wrong horizon segment. In each fault-bounded block, one or more additional seed picks are placed, and the algorithm tracks independently within each block. Where faults are numerous and closely spaced, as in the compressionally deformed structures of the Alberta Foothills or the extensionally faulted grabens of the North Sea, seeding every fault block manually can itself be time-consuming, and the quality of the final horizon map depends heavily on the interpreter's ability to correctly identify fault block boundaries from the vertical and time-slice displays before beginning the autotrack. Modern interpretation workflows use a combination of automated fault detection (coherence or semblance attribute volumes) to first define the fault network, then autotrack within each fault block, finally merging the block-by-block picks into a single continuous horizon surface with geologically correct fault-plane terminations.
• Amplitude extraction and attribute maps from autotracked horizons: Once a horizon is autotracked through a 3D seismic volume, the pick defines a set of time values (or depth values after depth conversion) at every inline-crossline grid node covered by the survey. From this horizon surface, interpretation software can extract a wide variety of seismic attributes along or near the horizon, creating 2D attribute maps that reveal lateral variations in reservoir properties. The most commonly extracted attributes are RMS amplitude (within a window above or below the horizon), maximum amplitude (the largest absolute value within the window), average absolute amplitude, instantaneous amplitude envelope at the horizon time, instantaneous phase, and instantaneous frequency. These attributes are used as DHI indicators, porosity proxies, and lithology discriminators across plays such as the Montney, Duvernay, and Viking in Alberta and BC. The accuracy of the attribute map is entirely dependent on the accuracy of the autotracked horizon: if the pick is offset by even one or two samples (2-4 ms) from the true reflection peak due to tracking instability in noisy areas, the extracted amplitude values will be systematically low relative to the true amplitude at that location. This is why autotracked horizons are always visually QC'd by the interpreter using vertical sections, time slices, and horizon variance maps before amplitude attributes are extracted for quantitative analysis.
• Stratigraphic autotracking and channel delineation: Beyond structural horizon picking, autotracking is applied to stratigraphic interpretation problems where the objective is not a regionally mappable reflector but a lithologically defined unit such as a channel, point bar, fan lobe, or reef edge. Stratigraphic autotracking typically uses a combination of amplitude and similarity criteria within a time-bounded stratigraphic window defined by two bounding horizons, searching for connected bodies of high amplitude or high coherence that represent the target facies. For example, in Cretaceous channel-sand plays of the WCSB (Nikanassin, Falher, Spirit River) the interpreter first autopicks the bounding shale reflectors above and below the channel interval, then runs a stratigraphic autotrack within the window using RMS amplitude as the tracking criterion, delineating the bright-amplitude channel bodies that correspond to clean, gas-bearing sand relative to the low-amplitude shale background. The result is a body of connected high-amplitude voxels that maps the channel geometry in three dimensions, which is then used to position horizontal wells along the channel axis and estimate the sand volume for reserve calculations. The accuracy of this approach depends on having sufficient resolution to distinguish the channel from its surroundings and on the amplitude contrast between the sand and the encasing shale being large enough to create a trackable amplitude anomaly.
• QC methods and tracking failure recognition: Autotracked horizons must be systematically quality-controlled before use in interpretation, depth conversion, or volumetric calculations. The primary QC tools are the horizon variance or tracking confidence attribute (low confidence = tracking instability = review required), vertical section display with the pick overlaid to confirm it follows the correct reflection event, time-slice or horizon-slice display to detect unphysical jitter or jumps in the picked time, and well-tie comparison where available wells provide ground truth for the two-way time at a known formation top. Common autotracking failures include cycle-skipping, where the algorithm jumps from one reflection peak to an adjacent peak at a different two-way time, producing a horizon map with discontinuous steps of half or full wavelet period (typically 10-25 ms); sidetrack into a noise burst or multiple, where the pick migrates off the geological reflection and tracks a noise event for some traces before losing continuity; and tolerance-too-wide errors, where the algorithm accepts picks that deviate significantly from the seed waveform in areas of poor signal, producing a geologically unrealistic horizon surface. Horizon variance maps, which display the root-mean-square of the time difference between each pick and its neighbors, are particularly effective at identifying spatially localized tracking instabilities that require manual correction or reseeding before the horizon is used for amplitude extraction or depth conversion.