Auto-tracking: when it wins, when it fails

How auto-tracking actually works

Strip away the marketing and production auto-trackers do something very simple:

• Read the seed pick's amplitude sign, call it positive (peak) or negative (trough).
• Move to the next crossline.
• Within the dip tolerance of the last pick, a small time window around it, search for a local extremum of the SAME sign; where there are several, take the strongest.
• If one is found, AND its amplitude is above a gate (a fraction of the seed's), AND the waveform around it still correlates with the seed's, place a pick there.
• Centre the next search on that pick and move on. The waveform is still compared with the original seed, not with the last pick, so a slow drift cannot quietly reset its own reference.
• Stop when any test fails (no extremum of the seed's sign within the dip tolerance, amplitude below the gate, a waveform that no longer matches) or at the edge of the survey. Those are the four reasons a tracker stops.

That's the whole algorithm. No machine learning, no pattern recognition in the modern sense. Its power comes from the fact that real reflectors are laterally continuous in amplitude, polarity, and time, so a 1-D continuity follower tracks them well. Its limits come from the cases where that continuity breaks.

Where auto-tracking wins

• High signal-to-noise zones. When the target reflector is an order of magnitude brighter than the noise, the extremum search finds the right event every time.
• Continuous, gently-dipping reflectors. Normal sedimentary reflectors with no tectonic complications are the bread-and-butter case. Auto-track thousands of traces in seconds, then QC the map for anomalies.
• Dense, regular survey geometry. When consecutive traces are close in space (10 m to 25 m bin size) the reflector barely moves between traces and the tracker's small search window succeeds.
• Zero-phase data. The peak sits on the reflector, so snap-to-peak behaves. Phase-rotated or minimum-phase data confuses the tracker because "the peak" may not be on the reflector.

Where auto-tracking fails

• Faults. The reflector on the other side of the fault sits at a different time. The tracker, which only knows about local continuity, either drops the pick (if no extremum of the right sign lies within the dip tolerance, or the one that does no longer matches the seed) or walks onto a DIFFERENT reflector on the other side of the fault that happens to be near the tracker's last time estimate. The second case is especially dangerous, the tracker keeps going, confidently picking the wrong reflector across the entire downthrown block.
• Stratigraphic pinchouts. As a bed thins, the reflections from its top and base interfere: the composite event first brightens, most strongly near a quarter of a wavelength (tuning), then fades as the bed thins further. The tracker's amplitude gate eventually rejects it and the horizon stops, often correctly, but in a way that can hide a pinchout that is still geologically real.
• Low signal-to-noise or attenuation zones. In deep, attenuated parts of the volume, the "brightest extremum" may be noise rather than the reflector. The tracker picks noise.
• Overlapping reflectors of similar brightness. When two reflectors are near each other and comparably bright, the tracker flips between them, producing a "noisy" horizon that is really a hybrid of two different surfaces.
• Lateral facies change. A sand grading into a shale changes the reflector's sign or strength. The tracker stops (polarity flip) or walks onto a different event.

Exercise

• Seed. Clear the horizon, set A click on (a) to Tracks, and click once on the target near 700 ms. Picks run both ways from the click, and the report names, for each side, the test that stopped it and the crossline. That line is the difference between watching a tracker stop and knowing why.
• Watch the confidence. Every tracked pick carries its waveform correlation with the seed, and the line in (a) fades as that correlation approaches the Waveform match threshold, where the tracker stops. A faded stretch warns that the next trace may be where it gives up.
• Propagate and grow. Press Propagate to next inline, which carries every pick to the next inline through the same four tests, then Auto-track this inline, which regrows every run outward until it meets a neighbour or stops. It never overwrites a pick you already have, and pressing it twice does nothing, because regrowing from its own frontier would reset the correlation reference and let the horizon walk onto another event. Ten rounds from the seed above give 1,104 picks on 11 inlines at an RMS of 1.2 ms, none on the wrong event, covering 30 % of the map; the gaps that stay open are real breaks.
• Loosen the tolerance at a fault. The second exercise seeds crossline 2106 on inline 1039 with a 28 ms dip tolerance. Across the fault near crossline 2057 the tracker steps 28 ms deeper where the reflector steps about 34 ms shallower, against the throw, onto a peak two cycles deep, finds it similar enough to the seed to pass the waveform test, and runs on to the edge: 115 picks, 57 of them on the wrong event, an RMS of 52 ms. At 16 ms the same seed stops at the fault with 58 picks at 1.1 ms. Dropping the pick is the safer outcome; the dangerous one looks exactly like success.
• Fading and reversing. The third exercise tracks two other reflectors on inline 1004. The pinchout top is seeded at tuning, where the wedge is 14 ms thick. On its thin side it stops at crossline 2109, where its amplitude has fallen to 18 % of the seed's, below the 28 % gate; on its thick side it stops at crossline 2070, where top and base have separated and the composite waveform no longer matches the seed (correlation 0.64). The facies boundary stops at crossline 2072, where its coefficient is weakening towards zero (it changes sign near crossline 2079) and the waveform no longer matches the seed (correlation 0.61). All four stops are right. Past each one the event must be seeded again by hand, and beyond the facies change it is a trough.
• Grade. Press Grade. The clean run above is Excellent; the loose-tolerance run is Check your work, and the written grade names the cycle skip.

The QC discipline

Auto-tracked picks are never final until they have been inspected. The time map is your primary QC tool. What to look for:

• Smooth colour gradients. Sedimentary reflectors produce maps that look like smooth topography. Sudden spots of very different colour in the middle of a smooth field are almost always pick errors.
• Sharp linear discontinuities. These may be faults (real) or tracker failures (not real). Cross-reference with your structural picks from Section 2.5 (when we build the fault picker) and with coherence attributes (Part 6). A lineament on the time map that has no corresponding fault on the structural interpretation is a bug, not a feature.
• Areas where picks drop out. The tracker gave up. Usually because the reflector became too dim or the dip too steep for the tolerance. You may need to re-seed in that area manually, or widen the tolerance and re-track.
• Near-well consistency. If you have wells, check that the auto-tracked horizon matches the well marker at each well within your expected depth uncertainty. Any systematic offset at wells means the tracker walked onto the wrong reflector somewhere between the seed and that well.

A working rhythm

The professional workflow combines manual and automatic:

• Seed manually at wells (highest-confidence anchor points).
• Seed manually at a few widely-spaced other inlines to form a skeleton.
• Auto-track from each seed, filling the volume.
• QC on the time map. Fix problem zones (delete bad picks, re-seed, re-track with adjusted tolerances).
• Cross-check at wells one more time.
• Ship the horizon to the downstream workflow (depth conversion, attribute extraction, volumetric calculations).

The step that junior interpreters skip is step 4, the QC. The interpreters whose maps get drilled through without nasty surprises are the ones who treat step 4 as the most important step, not the last one.