AI drone defect detection, geolocated on the model
Imagery analysed for cracking from 0.2 mm, spalling and corrosion staining, every detection checked by hand, then pinned to its position on the 3D model and issued as a geolocated PDF report your team can open on site.
AI defect detection runs automated vision over systematic, close-range drone imagery of a structure, flags cracking, spalling and corrosion staining, and pins each finding to a photogrammetric 3D model so it has a location, not just a photograph. Every detection is reviewed by hand before issue; what you receive is the verified set, as a geolocated PDF report with the 3D model and a CSV/GIS register behind it. The 0.2 mm figure comes from a controlled capability trial on a concrete culvert under the M1 at Tingalpa. Indicative pricing as at October 2026, exclusive of GST: a single structure with AI defect reporting starts in the low four figures.
What the model finds
The structure is photographed systematically by drone, and a model trained on infrastructure surface defects finds and maps cracking, spalling and corrosion staining across the whole capture. It picks up cracks from 0.2 mm, finer than a ground-level visual pass reliably catches, and every detection is verified by hand before issue.
Each detection is then projected onto the reconstructed model, so a crack has coordinates, a member and a chainage.
Why the PDF report matters as much as the model
A register that only exists as points on a 3D model needs the right software, and usually a laptop, before anyone can open it. So the deliverable is a geolocated PDF report: photos, severity, chainage and location plotted on plan, readable on a phone in a site shed with no CAD licence or model viewer required. The 3D model is there for anyone who wants to interrogate it further; the PDF is built to be opened first.
| Crack detection | FROM 0.2 MM |
|---|---|
| Detection classes | CRACKING · SPALLING · CORROSION STAINING |
| Verification | EVERY DETECTION, BY HAND, BEFORE ISSUE |
| Output | GEOLOCATED PDF REPORT + 3D MODEL + CSV/GIS REGISTER |
DETECTION CROPWhere the 0.2 mm comes from
A concrete culvert under the M1 at Tingalpa was captured at a fine ground sample distance to find out how small a crack the workflow would resolve. The answer was 0.2 mm. It was a capability trial, not a client job, and it's the source of every 0.2 mm figure on this site.
Two things the number does not mean. It is a detection threshold, not a width measurement: finding a 0.2 mm crack and measuring its width to 0.2 mm are different problems, and detection runs well ahead of measurement. Nor is it automatic on every capture. Resolution is set by standoff, lens and light, so the flight plan for each structure is designed to hold it. For context, Queensland's Structures Inspection Manual classes cracks of 0.1 to 0.3 mm as minor; 0.2 mm sits in the middle of that band.
How it works
- 01
Systematic capture
Full-coverage imaging at fixed standoff and resolution: every face, repeatable waypoints.
- 02
Reconstruction
Imagery is built into the 3D model that becomes the spatial index for every finding.
- 03
AI detection
Each frame is analysed for cracking from 0.2 mm, spalling and corrosion staining, and each detection is scored.
- 04
Geolocated register
Findings are pinned to the model and issued as a register your engineer reviews, filters and signs off.
Every detection is checked before you see it
Automated vision finds more than it should: on real infrastructure it routinely flags water staining as cracking, paint flaws as rust, shadows as spalling. That's deliberate, because a model tuned to miss nothing over-reports. So every flagged feature is reviewed by hand before the report is issued, the false positives are discarded, and severity, dimensions and recommended action are confirmed on each one that remains.
What it finds, and what no camera can
The detection classes are surface classes: cracking, spalling and corrosion staining, on concrete and steel the camera can see. Anything inside the member is out of reach: the depth of a crack, whether it's still active, hollow concrete behind an intact face and the state of the steel under the cover all need other tools. A core or ultrasonic testing answers depth, hammer sounding finds delamination, cover-meter and half-cell work covers the steel, and repeat gauge readings over time show whether a crack is moving. The decision to commission any of them belongs with your engineer.
What the register does is send those tests to the right places. The full reasoning, including why detection and measurement are different problems, is in the explainer on AI crack detection by drone.
Built to support the engineer's review
Severity classification, intervention decisions and sign-off remain with the responsible engineer. What changes is the quality of what they are signing off on: complete coverage, consistent resolution, and findings they can open at full detail on the model.
Because capture waypoints repeat exactly, each cycle can be compared directly with the last: the same face, the same standoff, a year apart. The flying itself is standard close visual inspection work, with detection run over the imagery afterwards, and that flying has already been done on real structures, at the Indooroopilly Rail Bridge and the Story Bridge.
It suits concrete and steel structures where surface condition drives maintenance: bridges, dams and spillways, towers, wharves, facades and industrial plant. If reconstruction can see the surface, the pipeline can index it.
VISUAL INSPECTION — CORROSIONQuestions we get at scoping
Q/01How small a defect can the AI detect?
Cracks from 0.2 mm width, given capture at the designed standoff and resolution; that figure was established in a controlled trial on a concrete culvert under the M1 at Tingalpa. Every automated finding remains reviewable at full image resolution, so your engineer can check each one directly.
Q/02Does it measure crack width, or only find cracks?
It finds and maps them first. Each finding's extent and location are checked during manual review, but a crack detected at 0.2 mm isn't measured to 0.2 mm: measuring width reliably takes several pixels across the crack. Where a precise width matters, such as a crack being monitored, the register shows where a gauge reading is worth taking.
Q/03How are detections checked, and what gets thrown out?
Every flagged feature is reviewed by hand before the report is issued. Water staining read as cracking, paint flaws read as rust and shadows read as spalling are discarded; what remains is confirmed against the full-resolution frames. You receive the verified set, never the raw output.
Q/04Can it find delamination or corrosion inside the concrete?
No. It sees surfaces. Delamination needs sounding, reinforcement corrosion needs half-cell readings or a break-out, and crack depth needs a core or ultrasonic testing. The register's job is to show where those tests are worth doing.
Q/05Does this replace an engineering inspection?
No. It replaces the data-gathering: the access hire, the working-at-height exposure, the sampling compromises. Assessment, severity classification and sign-off stay with the responsible engineer. The register is structured to make that review faster.
Q/06Do I get a model file, or something I can just open?
Both. The deliverable is a geolocated PDF report alongside the underlying 3D model and a CSV/GIS register for anyone who wants to work with the data directly. The PDF is built to be opened first; the model is there when someone needs to measure.
Q/07How much does AI defect detection cost?
It's scoped on the asset: size, access and how much surface needs analysing. As a guide, a single-structure inspection with AI defect reporting starts in the low four figures, exclusive of GST. See the drone survey cost guide for the full picture; Hover returns a written, fixed quote within one business day.
Q/08Do we need AI defect detection, or is a visual inspection enough?
A close visual inspection is the more cost-effective option when an itemised register read by an inspector is enough: condition checks, verification jobs, routine cycles. Step up to AI defect detection when you need automated detection across the whole capture, with every finding pinned to a measurable 3D model. That's the tool for condition baselines, deterioration tracking, and large surface areas where reviewing every frame by eye isn't practical.
Scope this for your site
Send the asset and the deliverable you need. A scoped method and quote come back within one business day.
HARRY BILLING — CHIEF REMOTE PILOT
0411 445 458
[email protected]
CASA ReOC #9398 · $28M PUBLIC LIABILITY