Using AI & Machine Learning for Pipeline Geohazard Management

Pipeline geohazard programs have access to more information than ever. Lidar, InSAR, IMU, instrumentation, satellite imagery and field inspections can provide unprecedented visibility into changing ground conditions.

But for many mature programs, the problem is no longer collecting enough data. It’s interpreting all of it consistently and at scale.

As monitoring has expanded, many programs have become interpretation-limited rather than data-limited. Engineers need to find the relatively small number of signals that indicate credible integrity concerns, while maintaining consistency across large pipeline systems and focusing limited resources where they matter most.

That is where machine learning and AI can be particularly useful. Not by replacing engineering judgment, but by helping engineers know where to look.

AI as a triage layer

Machine learning is well suited to repeatable tasks involving large volumes of data: recognizing patterns, identifying anomalies and prioritizing locations for further review. Three applications developed by Cambio Earth and BGC Engineering illustrate how this can work in practice. In all three examples we use automation to handle scale, to focus engineering expertise where it adds the most value.

1. Screening IMU data for landslide interaction

Vendor bending-strain reports contain large numbers of features, the vast majority of them benign construction artifacts such as roped overbends and road bore tie-ins. Conventional threshold-based screening means reviewing roughly 20 non-landslide features for every landslide feature. A deep-learning model trained on 30,000 labelled bending-strain features, including 478 SME-confirmed landslide impacts, looks for deformation patterns consistent with ground movement. On a spatially distinct hold-out test set it achieved approximately 90% recall and 95% specificity.

Operationally, it reduces the features requiring SME review by ~5X compared with traditional screening criteria, or as many as 50 fewer analyst hours per average 110 km ILI run.

2. Mapping active landslides from lidar change detection

Lidar change detection can reveal ground movement down to 0.1 m, and works in forested terrain where vegetation obscures deformation in satellite imagery. But manual interpretation at corridor scale is time-consuming and varies between analysts. A computer-vision model was developed to segment candidate zones of active movement for expert review. The result surprised us: the team reviewed all 3,100 of the model's apparent false positives, and over 1,800 turned out to be real landslides that the original manual mapping had missed. Adding them back to the training data brought the final detection rate to 92.5% of landslides, within the Appalachian region the model was trained and tested in, primarily on relatively small earth slides. Work is underway to expand and validate the model in other geographies and landslide types, including Western Canada and North Carolina

The broader lesson is that the most complete inventory comes from ML followed by SME review, not from either alone.

3. Prioritizing watercourse crossings

Pipeline systems can contain thousands of watercourse crossings, roughly one every three kilometres, but risk is concentrated at very few of them. Our earlier work found that about 4% of hydrotechnical sites account for approximately 90% of total system-wide probability of failure. Field-inspecting all of them is disproportionate to the risk.

A probabilistic ML model trained on approximately 20,000 field inspections across North America estimates screening-level PoF from desktop-available data. Rather than predicting PoF directly, it predicts the physically meaningful components (depth of cover, the scour depth curve, vulnerability), which preserves interpretability, and it predicts full distributions rather than point estimates so conservatism can be tuned to risk tolerance.

In one case study, 1,603 crossings were screened and prioritized, reducing field inspection costs by an estimated 84%, or roughly $2M CAD.

What makes these models useful in engineering decision-making?

Good AI isn't just about the model. The training data needs to be high quality and representative. The model needs to be tested independently and within a clearly defined domain of use. Uncertainty needs to be understood. And performance should be measured against the engineering process the model is intended to improve.

Most importantly, the output needs to lead somewhere. If an algorithm identifies something interesting but that result isn't connected to engineering review, additional monitoring, field investigation or mitigation, it has simply created another data stream to manage.

The real opportunity is to integrate ML into the geohazard management program itself: continually screening new information, identifying meaningful changes and directing engineering attention toward the sites most likely to require action.

Viewed through a program maturity lens, ML is best understood as an accelerator between stages rather than a capability in itself. Landslide mapping from lidar change detection can move a program from ad hoc to inventory-based. Watercourse PoF screening can take an inventory-based hydrotechnical program to systemic and risk-informed. IMU screening, applied as a complete workflow, can move a program further still.

AI doesn't replace engineering expertise. It helps organizations apply that expertise consistently across increasingly large and data-rich programs.

Learn more at IPC 2026

Our own Sarah Newton, Chief Product Officer at Cambio Earth, will present “Applications of Machine Learning and Artificial Intelligence to Pipeline Geohazard Integrity Management Programs” at the 2026 International Pipeline Conference (IPC) in Calgary on September 25 during the 10:30 am-12 pm session. This paper was co-authored by Sarah Newton, Aron Zahradka and Andrew Johnson (Cambio Earth), Corey Scheip (BGC Engineering) and Doug Dewar (Pembina).

As well, Aron Zahradka, Data Scientist Lead at Cambio Earth, will present “Use of Machine Learning to Identify Landslide Signatures in IMU Bending Strain Data” at IPC on September 22 during the 10:30 am-12 pm session.

‍