Gains:
- Ability to understand the physics and limits of ERT, GPR and seismic tomography and choose the appropriate method for the target.
- Ability to verify GPR depths with velocity calibration and cross-interpret resistivity/velocity anomaly rather than using a single method
- Ability to consider the heterogeneity and seasonal effects of the near surface and confirm each anomaly with ground verification (drilling/trench).
Some geophysical problems are deep and large-scale; But some are very close at hand: a leak under a dam, a gap under a road, an archaeological structure, a shallow water table, or a pollution plume. Near-surface geophysics and underground imaging is the task of imaging the first few meters to hundreds of meters with high resolution. This unit will cover three common methods—electrical resistivity tomography (ERT), ground penetrating radar (GPR), and seismic tomography—and how artificial intelligence (AI) is accelerating the data processing, inversion, and interpretation of these methods; but we will see why the inherent uncertainties of the near surface require using AI with caution.
Three methods, three physics
Short dictionary. ERT (electrical resistivity tomography) is a method that extracts the underground resistivity (resistance to electric current) distribution by applying current to the ground and measuring the potential difference; It is powerful in separating water, clay and voids. GPR (ground penetrating radar) sends high-frequency electromagnetic waves and images shallow layers and objects (pipes, cavities, reinforcement) from reflections; It has high solubility but limited depth and fades quickly in conductive soil (clay, salt water). Seismic tomography extracts the subsurface velocity distribution from wave arrival times in many source-receiver paths; It is applied through crosshole or surface alignment.
What all three have in common is that they all produce an inversion: estimating the distribution of the subsurface property (resistivity, velocity) from measured data. And they all suffer from the same uncertainty: resolution decreases with depth, data is limited, and the result is regularization dependent.
Contribution of AI
AI shines in these methods in a few places:
- GPR interpretation: Automatically recognize hyperbola patterns (characteristic reflection given by buried point objects - pipe, rock) on the radargram (GPR section) and suggest location and depth.
- Accelerating ERT/tomography inversion: Shortening the iterative solution with the surrogate models we saw in the previous unit.
- Noise and interference removal: marking outlier measurements from antenna ringing in GPR, poor electrode contact in ERT.
- Automatic anomaly marking: Suggesting void, leak path, buried structure candidates throughout the volume.
The risks are also familiar: the hyperbola detector may mismatch reflections; the inversion accelerator can fill the deep region invisible to the data with the precursor; The anomaly detector may mistake ground heterogeneity for "structure".
Hint: In GPR, the curvature of a hyperbola gives the wave speed of the medium; The depth estimated by the AI depends on this speed. If the speed is wrong, the depth is also wrong. Do not present exact depths without calibrating the velocity on at least one known target (e.g. a pipe of known depth).
Special uncertainties of the near surface
Resolution is high at the near surface, but the ground is very heterogeneous: moisture, fill, roots, debris change with every meter. This is full of pitfalls for AI. In ERT, a low resistivity zone can be either water, clay, or a seep — resistivity alone does not distinguish. Therefore, near-surface ground truth (borehole, trench, known infrastructure) is critical; Every anomaly flagged by the AI is confirmed by a physical check, if possible.
Caution: GPR and ERT results are extremely sensitive to measurement geometry and ground conditions; The same line appears differently in the dry season and after rain. Don't make the AI say "this anomaly is definitely a gap"; Read the result with the context of measurement condition, season and alignment.
Multi-method near-surface interpretation
One method is often not enough on the near surface. ERT gives resistivity, GPR gives structure, seismic tomography gives speed; All three combined together lead to a much safer interpretation of low resistivity + high amplitude GPR reflection + low velocity, such as a "water-filled space". AI helps combine these layers (data fusion), but under expert supervision that respects the resolution and depth limits of each method.
three mini cases
Case 1 — Dam leak path. An earthfill dam was suspected to be leaking. A low resistivity band was seen with ERT; The AI anomaly detector flagged this as a leak path candidate. The team confirmed it with GPR and two soundings; The band was truly a saturated transition zone. If the decision was made by only one method, it could be confused with a clay lens.
Case 2 — Hyperbola fallacy. In an infrastructure scan, the AI marked 14 hyperbolas and said "14 pipes". The calibration showed that the velocity was incorrect: the actual depths were 30% shallow and the two hyperbolas were actually multiple reflections of a single pipe. When the speed was calibrated from a known line, the number dropped to 11 and the depths improved.
Case 3 — Seasonal effect. At one archaeological site, ERT gave "no structures" at the end of summer. Repeated measurement after the rain revealed wall foundations: contrast was low on dry ground. If the AI interpretation had relied on the initial data and declared the field "empty", the structure would have been missed.
Four copyable templates
1) Method selection consultation:
Your role: near-surface geophysicist. My target: [buried cavity / seepage / archaeological structure / water table], soil: [clay/sand], depth: [0-5 m]. Which one(s) of ERT, GPR and seismic tomography is suitable and why? Explain the strengths, limitations, and complementary uses of each method for this goal.
2) GPR speed calibration:
There are hyperbola patterns on my GPR radargram. How do I estimate the ambient speed from the hyperbola curvature, how do I calibrate it to a known target? Explain the effect of wrong speed on depth error and how to distinguish multiple reflections from the real object.
3) ERT anomaly discrimination:
There is a region of low resistivity in my ERT inversion. How do I separate the possibilities of this being water, clay or space? Which data (GPR, drilling, seasonal repetition) clarifies the distinction? Explain why resistivity alone is not sufficient.
4) Ground verification plan:
AI flagged several anomaly candidates in volume. Propose a ground verification (drilling/trench) plan to physically verify these: which anomaly do I test first, how many points are needed at least, how do I compare the result to the AI map?
Weak prompt / Strong prompt
Weak prompt:
Show me if there are any gaps in this GPR data.
No speed calibration, no ground context and no verification; The AI counts hyperbolas, but the depth and type are unreliable.
Powerful prompt:
Your role: GPR expert. Context: concrete subway, target cavity/pipe, ground damp, known pipe [X] m depth. Task: (1)calibrate the medium velocity through the known pipe; (2) mark the hyperbolic patterns and calibrate the depth of each; (3) distinguish multiple reflection from real object; (4) classify the island as cavity or pipe with confidence level and suggest ground verification. Don't use strong language.
Calibration, type discrimination and verification request make the output reliable.
Method comparison
feature
ERT
GPR
seismic tomography
measured by
resistivity
EM reflection
wave speed
is strong
water/clay/space
Shallow structure/object
speed structure
weakness
low resolution
conductive ground
road coverage
AI contribution
Inversion speed
Hyperbola recognition
Inversion speed
Common mistakes
- Giving GPR depth without calibrating the speed. Wrong speed means wrong depth.
- Interpreting resistivity alone. Low resistivity can be water, clay or space.
- Considering multiple reflections as real objects. Watch out for repeating patterns in GPR.
- Ignoring the season/condition effect. Dry/wet ground gives completely different results.
- Bypassing ground verification. Near surface drilling/trench confirmation is critical.
In summary
ERT, GPR and seismic tomography image the subsurface with high resolution; AI makes a strong contribution to hyperbola recognition, inversion acceleration, and anomaly marking in GPR. But the extreme heterogeneity of the near surface, the resistivity/velocity uncertainty, the need for velocity calibration, and the seasonal effect make the AI output fragile. Correct use; calibrate the rate, combine methods, do not interpret resistivity in isolation, and confirm each anomaly with ground verification. The final interpretation lies with the expert who knows the resolution limits.
Application task
Select a near surface target (cavity, leak, buried structure). Determine the appropriate method(s) with the "Method selection consultation" template. If you chose GPR, plan depth reliability with the "GPR speed calibration" template; If you chose ERT, set up the water/clay/space separation with the "ERT anomaly separation" template. Write a ground verification plan for at least one anomaly.
checklist
- [ ] I chose the appropriate method(s) with justification.
- [ ] I calibrated the GPR depths with the known target.
- [ ] I have cross-interpreted the resistivity/velocity anomaly rather than using a single method.
- [ ] I took into consideration the effect of season/measurement condition.
- [ ] I confirmed the anomalies with the ground verification plan.