CASE 13 · CASE ATLAS · LEVEL B WORKSHOP SCENARIO

Offshore foundation corrosion survey

SITUATION

An annual structural survey of an offshore wind foundation produces thousands of drone and ROV images of the splash zone and subsea structure. The inspection lead needs the images most likely to show coating breakdown or structural damage surfaced first, without letting a missed frame become a missed finding.

EVIDENCE

The full image set per foundation with position and depth metadata, the coating and structural defect classification criteria, and a qualified surveyor's prior dispositions on a labelled sample of images.

AI APPROACH

An image-based detection model proposes candidate coating-breakdown and damage locations with a confidence estimate, presented to the surveyor as a ranked review queue with the source image and position, not as a pass/fail structural verdict.

KEY CONTROLS

The model's coverage is stated explicitly — what fraction of the image set it processed and at what confidence threshold — so an unprocessed or low-confidence frame is a visible gap, not a silent pass; no finding is closed without a qualified surveyor's sign-off.

VALIDATION MEASURES

An attribute agreement study compares the model's flags against surveyor consensus on known-defect and known-clean images from a prior survey, tracking missed defects and false flags separately given their different safety consequences.

HUMAN DECISION

The qualified surveyor accepts, downgrades or escalates every flagged finding and signs the structural survey report; the model narrows what gets human attention first, it does not close the survey.

DISCUSSION QUESTIONS

  1. What happens to this year's survey conclusion if 5% of frames were too turbid or poorly lit for the model to score?
  2. Which is worse for this structure — a missed early-stage coating breakdown or a flagged non-issue — and does the review order reflect that?
  3. How does this model get re-validated after a foundation type or camera rig changes?

TRY IT LIVE

Run this scenario's approach through Attribute Agreement Study — Do reviewers and the system agree on the same cases?

Open Attribute Agreement Study →