AI pothole detection: what councils need before choosing a repair

1. Define the decision the system will support

Write the operational question in one sentence. For example: “Which reported defects need inspection before we plan next week’s maintenance?” That is a different job from estimating repair quantities, prioritising a network or recommending a treatment. A pilot attempting all four at once is difficult to evaluate because an error can enter at several stages.

Record the current process as a baseline. Who receives a report, checks the location, inspects the surface and approves the work? Where is information missing or being entered twice? The first useful improvement may be a reliable evidence record that reaches the right person, rather than an automated repair recommendation.

Agree what the tool is allowed to suggest and what requires an assessor. Keep a visible distinction between something observed in a photograph, something calculated by the system and something approved after review.

2. A pothole and a repair area are different things

NSW’s 2023 Regional Emergency Road Repair Fund guidelines distinguish filling potholes from heavy patching, which removes and replaces failed material or stabilises it. They also describe rehabilitation where failure extends beyond practical heavy patching. These are useful examples of different maintenance categories, not a current universal treatment specification or an open funding offer.

MidCoast Council explains why surface appearance is incomplete evidence: roads that look similar can have different conditions underneath. Water, drainage and the state of the pavement affect the maintenance needed. The council also distinguishes temporary patching to address an immediate hazard from later permanent work.

For a buyer, the implication is straightforward: ask how the system exposes uncertainty and missing context. A neat outline around a hole does not establish the full area that should be repaired. The responsible road team must determine the investigation and treatment appropriate to the site.

3. Specify a useful defect record

Use the following as a starting point for a procurement brief. Each field should have a clear origin and a way to correct mistakes. An estimated measurement should identify its method and uncertainty; an ordinary image without a reliable scale cannot substantiate precise dimensions.

Location matters beyond putting a dot on a map. Service NSW asks reporters to note where damage is located and explains that reports on local roads can be passed to the relevant authority. Your own workflow should identify the road owner and person responsible for assessment rather than relying on an assumed boundary.

4. Test a decision with two similar-looking defects

Consider a fictional example. Two images show similarly sized holes. One is an isolated defect with an intact surrounding surface. The other is beside repeated repairs and a wider area of visible distress. Neither image alone proves what lies underneath. The second record, however, gives an assessor a reason to seek more context before accepting a narrowly drawn repair area.

A useful demonstration lets the reviewer compare those records, open the original evidence and record a different assessment. A weak demonstration presents one confident treatment label and hides how it was produced. Ask the supplier to demonstrate a disagreement, a missing image and an uncertain location as well as a successful detection.

This is an evaluation exercise, not an engineering instruction. Qualified road personnel should set the investigation and treatment rules for the roads they manage.

5. Measure the whole workflow

Build the test set with your road team. Include different surfaces, lighting, weather, capture equipment and defect types within the intended scope. Reserve a final set that was not used to tune the system. Compare results with reviewed records and document where assessors disagree.

Report missed defects, incorrect detections and duplicate reports separately. Evaluate measurements against an appropriate reference where quantities are claimed. Then check the work process: can staff locate the finding, understand it, correct it and assign the next step? Count review time including corrections, not just model processing time.

Do not combine everything into one accuracy percentage. A duplicate can waste inspection time; an incorrect location can send a crew to the wrong place; a missed hazardous defect has a different consequence. Agree which failures are unacceptable and when the pilot should pause.

6. Connect the pilot to the people doing the work

Before procurement, agree the output format and how records enter the existing asset or work-order system. Decide who can amend a finding, how earlier versions remain visible and what happens when a model changes. Ask for an export you can inspect without the supplier’s application.

DataXLR8 has delivered a pothole estimation and repair-scoping engine for Transport for NSW. Our public road demonstration illustrates the problem using synthetic geometry. It does not reproduce client inspection data or claim measured project results. We can discuss the requirements for a similar system and the evidence needed to evaluate it.

Start with one bounded question, a representative sample and a review session with the people who will use the output. That gives both the buyer and delivery team a concrete basis for deciding whether to expand the work.