Can You Auto-Detect Landslides from IMU Data? Now You Can.

Historically, pipeline operators used IMU data primarily to verify the position of their pipelines. IMU data can also be used to generate a bending strain report, but understanding what the results mean takes significant specialist time and effort. As a result, many reports are difficult to review and action, and even when they are, typically only a small number of strain features receive full review. But what if you could auto-detect landslides from the same data with minimal effort?  

This is an early preview of a new way to turn IMU data into actionable geohazard insight—built for pipeline integrity teams.

Beyond Peak Strain: Interpreting the Full Deformation Pattern

Experienced geohazard engineers don’t just look at peak strain values. They recognize patterns:

• Broad, asymmetric horizontal strain lobes
• Spatial context across tens to hundreds of meters
• Subtle signals that only make sense when viewed as a whole

That pattern recognition is hard to codify as rules, but that’s exactly the kind of task machine learning is good at.

So we asked a simple question: Can a model learn to recognize landslide signatures in IMU data the way a geotechnical SME does?

A New Way Forward: Cambio IMU Assessment (Preview)

The paper being presented at PPIM 2026 describes a convolutional neural network (CNN) trained to determine the likelihood that a bending-strain feature is landslide-related.

Key points, in integrity-team terms:

• Input: Raw IMU-derived horizontal strain, vertical strain, pitch, and heading
• Context window: 200 m per feature (not just a single peak)
• Training data: >30,000 strain features from real transmission pipelines
• Ground truth: SME-labeled landslide assessments supported by lidar, field inspections, and historical monitoring

This is not a theoretical model trained on synthetic data. It’s trained on the same messy, vendor-variable IMU data integrity teams already deal with every day.

How It Performs (Compared to Today’s Screening)

To keep the comparison fair, the model was benchmarked against a common industry screening approach:

• Horizontal strain ≥ 0.125% OR
• Total strain ≥ 0.2%

Both methods were tuned to catch ~90% of known landslide impacts. This allows for a like-for-like comparison focused on screening efficiency rather than risk tolerance.

Here’s where things get interesting:

• Traditional screening: ~21 non-landslide features identified for every real landslide
• ML screening: ~4 non-landslide features identified for every real landslide

That’s an 83% reduction in false positives at the same detection rate.

In practical terms:

• Less analyst fatigue
• Faster identification of real threats
• More consistent triage across runs, vendors, and teams

And because the model outputs a probability, integrity teams can tune it to match available review capacity — instead of hardcoding a single strain threshold forever.

What This Is — and What It Isn’t

Let’s be very clear.

This is an early preview, not a magic black box.

It is: 

• A screening and prioritization tool 
• Designed to assist, not replace, expert judgment 
• Optimized initially for Appalachian transmission pipelines

It is not: 

• A replacement for geotechnical interpretation
• A substitute for lidar, field inspections, or multi-run analysis

Think of it as a very fast, very consistent screening analyst who never gets tired — and always shows their work for you to verify.

Why This Matters for Integrity Programs

Integrity teams are being asked to do more with the same data:

• Demonstrate defensible, auditable decisions
• Respond faster to emerging geohazards
• Justify why one site matters more than another

Automated IMU screening changes the conversation.

Instead of: “We flagged these because they exceed a threshold”

You get: “These features exhibit deformation patterns consistent with known landslide behavior”

That’s a meaningful shift toward science-first, risk-informed integrity management.

What Comes Next

This first release focuses on single-run IMU data. Future work is already underway to:

• Expand training data beyond Appalachia
• Incorporate run-to-run strain changes
• Integrate contextual data (terrain, lidar change, landslide inventories)
• Improve interpretability and QA/QC workflows

In other words: this is the first step, not the finish line.

Joint Research, Real Data

This work is the result of a close collaboration between Cambio Earth, BGC Engineering, and multiple pipeline operators combining:

• Real-world IMU datasets from operating transmission pipelines
• Decades of geohazard and IMU interpretation experience
• A data platform designed specifically for geohazard integrity programs

The research is being formally presented at the Pipeline Pigging and Integrity Management Conference (PPIM) and is being released here as an early preview for the industry. 

How Can You Get Involved?

If you manage IMU data, geohazards, or integrity risk — this work was built for you.

Because IMU data shouldn’t just be archived.

It should tell you where the ground is moving — before the pipeline does.

Integrity Manager FAQ

Is this replacing our geohazard engineers or IMU reviewers?
No. This is a screening and prioritization tool, not a decision-maker. The intent is to reduce noise and let your geotechnical and integrity SMEs spend their time on the features that actually look like landslide interaction. Final interpretation and decisions remain with qualified engineers.

How is this different from existing strain thresholds?
Traditional thresholds look only at magnitude. This model evaluates the shape, symmetry, and spatial context of the strain signal across a 200 m window — closer to how experienced analysts actually interpret IMU plots.

Can this be defended internally and with regulators?
Yes — because it does not operate in isolation. The model output is a probability used for triage, and every flagged feature is still reviewed using standard IMU plots, lidar, and other contextual data. The process is traceable, repeatable, and consistent across vendors and runs.

What data does the model use?
Only data most operators already have: raw IMU-derived bending strain, pitch, and heading. No new field data collection is required.

Where does it work today?
The early preview is trained primarily on transmission pipelines in regions with well-documented landslide activity. Performance will continue to improve as additional regional data are incorporated.

How would this fit into our existing integrity workflow?
The model runs immediately after IMU ingestion, ranks strain features by landslide likelihood, and integrates with existing geohazard and integrity management workflows. Teams can tune probability thresholds to match available review capacity.

‍