AI Applications · Energy and Facility Inspection

Change Detection along Pipeline Corridors

The useful question along a pipeline corridor is not what is there, but what changed since last time. Repeat survey with matched geometry turns thousands of images into a short list of changed locations.

5 of 10Method note

Overview

Comparability comes from matched flight lines, altitude and imaging parameters — a different geometry produces differences that are not real change.
Seasonal vegetation, lighting and surface moisture generate large volumes of apparent change; filtering them is most of the work.
Detected change is not automatically a risk: third-party excavation, subsidence, encroachment and legitimate maintenance require different responses.
Output is a location with a polygon and imagery pair, delivered into the client's GIS or work order system rather than held in the detection tool.

Change detection narrows the review workload. Whether a change is a threat is decided by the operator.

Corridors are managed by exception, not by coverage

A pipeline operator already knows roughly what the corridor contains. What they need is the difference between this month and last month — the excavation that started on Tuesday, the bank that slumped after rain, the structure that appeared inside the easement. Coverage is not the problem; timely exception reporting is.

This framing matters because it changes what the aircraft has to do. It is not carrying out an inspection in the usual sense. It is producing a repeatable observation — the same lines, the same height, the same imaging parameters — so that differences between two passes mean something physical rather than something geometric.

Comparability is the whole problem

Change detection compares two raster sets. If flight lines differ by a few metres, or if one pass was flown in flat light and the other in hard sun, the comparison produces differences everywhere — misaligned edges, shifted shadows, changed pixel values across whole tiles. These are not real changes, and they will bury the handful that are.

So the flight plan is a specification, not a suggestion: line placement, altitude, overlap, camera settings, and ideally a similar time of day and sun position. Seasons matter too. Comparing leaf-on imagery against leaf-off imagery along a wooded corridor will produce a wall of apparent change that no amount of threshold tuning will cleanly remove.

Registration and orthorectification have to be handled before detection, not patched afterwards. Sub-pixel alignment differences along thousands of frames are invisible one image at a time and overwhelming in aggregate.

Most detected change is benign

A difference is a candidate, not a threat. Third-party excavation, ground subsidence, unauthorised construction, vegetation encroachment, flooding, and perfectly legitimate maintenance activity all produce polygons that look similar in the first pass of analysis. Distinguishing them relies on context the imagery alone may not carry — work permits, notifications, seasonal work patterns.

This is why the output should be delivered with its context: change polygon, before and after image pair, location along the corridor, and whatever ancillary data the operator already holds. Reviewing that bundle is fast; reviewing a bare list of polygons is not.

Routing into existing systems is described in linking asset registers to imagery — the same discipline here determines whether the finding reaches a duty officer or sits in an analyst's folder.

Where this does not help

Very slow change is poorly served by differencing. Gradual movement that accumulates over years may fall below the detection threshold every single month and then surface as a large single event. For those mechanisms, dedicated monitoring rather than periodic differencing is the appropriate instrument.

Change detection also cannot tell you what something is for. It can say that a rectangular area became bare earth between two surveys. Whether that is an authorised laydown yard or an encroachment is an operator judgement supported by records.

Change Detection along Pipeline Corridors

Overview

Corridors are managed by exception, not by coverage

Comparability is the whole problem

Most detected change is benign

Where this does not help

Related pages

More in this area

Tell us what the system needs to do