Capstone: Ultra-high-frequency near-surface imaging

The processing of Parts 0 to 9 scales from exploration, with targets of kilometres to hundreds of metres, down to engineering geotechnics, provided the time and space sampling, the noise suppression and the velocity model all follow the band upward by an order of magnitude. This capstone walks a typical near-surface survey: locating voids, buried pipelines and shallow faulting in the top 30 m with a hammer or weight-drop source whose usable band reaches several hundred hertz.

Project setup

A geotechnical survey before an infrastructure project. A 500 m 2D line, 0.5 m geophone spacing (100 Hz geophones on spikes), an accelerated weight-drop source with a usable band of about 60 to 800 Hz. Sample interval 0.125 ms (Nyquist 4 kHz), record length 200 ms. Targets: voids 2 to 5 m across, pipes 0.3 to 1 m in diameter at 1 to 15 m depth, faults with 0.5 to 2 m of throw. A low signal-to-noise environment: traffic, construction and machinery nearby.

The pipeline

In the figure the void, centred 7 m down in the clay, is lost on the raw shots: at its place the section stands only 3.6 dB above the background, and we call it visible at 6 dB. Through the full flow, with dense first breaks, noise attenuation, deconvolution, inverse $Q$ and a migration with the refraction velocities, it stands 10.0 dB above the background, and its image sits 5.5 m deep, half a metre above the void's top at 6 m. The inverse-$Q$ stage does not pay here: with it left out the migrated void stands 11.4 dB above the background. At a 2 ms sample interval the recorder's anti-alias filter cuts into the 400 Hz band, and the void is lost again.

Differences from exploration-scale seismic

• Bandwidth. About 60 to 800 Hz near the surface, shrinking with depth (against 5 to 100 Hz for marine exploration). Vertical resolution is the tuning thickness $v/(4 f\_{\\mathrm{dom}})$: about 0.5 m for $v = 1500$ m/s and $f\_{\\mathrm{dom}} = 750$ Hz.
• Data volume. Each shot is at most 200 ms at 0.125 ms, 1600 samples: small per shot, but a dense line has about 1000 shots, and high throughput is expected.
• Q is the enemy. Attenuation costs $8.686\\,\\pi f t / Q$ dB: at $f = 3$ kHz and $Q = 50$ that is about 24 dB after 15 ms and about 160 dB after 100 ms, far below any noise floor. The highest frequencies simply do not survive to depth, so inverse-$Q$ compensation (Section 7.3) must be gain-limited to the band still above the noise.
• The near surface is the target. Refraction tomography is a primary product here, not just a statics correction, so every picked first break matters.

Why ML matters more near the surface

ML pays off more in near-surface work than in exploration because:

• Noise is heavy and variable: coherent ground roll and the 330 m/s air wave, plus bursty traffic and machinery noise. f-k filters remove the linear events but smear the bursts, and f-x deconvolution handles only the random part; a CNN trained on local noise can remove both.
• First-break pick counts are huge (every trace in every shot); ML automation saves days of interpreter time per project.
• Training data exists: geotechnical contractors hold years of reference surveys, and synthetic shot gathers are cheap to generate for pre-training.

PSDM at metre scale

Post-stack time migration is inadequate for metre-scale targets: the near surface is strongly heterogeneous laterally, and time-domain velocities smooth over exactly the features we want to resolve. Pre-stack depth migration on a 0.25 to 0.5 m grid collapses each void's diffraction hyperbola into a bright, polarity-reversed focus at its true depth, with an amplitude shadow beneath it. The figure stands in for it with a zero-offset diffraction summation using the refraction velocities; leave the first breaks out and its 1200 m/s guess, far faster than the soil and clay, overshoots the focus (the exercise Migrate without the refraction velocities).

Where this goes next

Section 10.6 returns to 4D processing with the most stressful case: comparing 2006 legacy streamer data with 2024 OBN monitor data. Heterogeneous acquisitions force every technique of Part 8 to work before the difference is useful.

References

• Yilmaz, Ö. (2001). Seismic Data Analysis (2 vols.). SEG.
• Claerbout, J. F. (1976). Fundamentals of Geophysical Data Processing. McGraw-Hill.
• Baker, G. S. (1999). Processing Near-Surface Seismic-Reflection Data: A Primer. SEG.
• Steeples, D. W., Miller, R. D. (1998). Avoiding pitfalls in shallow seismic reflection surveys. Geophysics, 63, 1213.
• Sheriff, R. E., Geldart, L. P. (1995). Exploration Seismology (2nd ed.). Cambridge UP.