AI-generated image, for illustration purposes only.
When the human eye can no longer keep up
Is that a leak or a shadow?
Someone asked it while reviewing images a drone had captured over a stretch of pipeline. From the air everything looked in order, except for that stain that didn't quite add up. It could be hydrocarbon seeping out. Or it could be the reflection of a cloud on the ground.
That doubt, repeated thousands of times over a year of operation, was the starting point of this project.
The problem wasn't the technology. It was the scale.
The client had been flying drones over its oil fields for some time, to detect incidents and monitor the most sensitive areas of its infrastructure. The hardware investment was made and it worked. The bottleneck showed up somewhere else, where almost no one looks at first: in what happens after the drone lands.
A single inspection campaign over one stretch of pipeline could generate more than 2,000 images in a day.
And they all ended up depending on the same thing: a person, sitting in front of a screen, going through image after image, looking for that stain that doesn't add up. The method works when there are few flights. But the operation grew, and the volume started arriving faster than any team could review. Images piled up, and the time between capturing a possible leak and seeing it stretched from hours to days.
There was an even less visible problem. When monitoring depends solely on the human eye, it also depends on fatigue, on the time available, and on how much focus that person has left at the end of the shift. The eighth reviewer of the day doesn't look the same way as the first.
In Oil & Gas that's not an operational detail. A spill detected late means environmental damage, regulatory exposure, fines, production stoppages, and a remediation cost that scales with every hour that passes. The risk was that the very success of aerial inspection would make it unmanageable.
We changed the question
Instead of asking how to review more images, we started asking how to find what mattered sooner. It's a small change in the wording and an enormous one in its consequences: you stop looking for more hands to watch screens and start looking for a way for the screen to alert you.
The answer was a prototype based on pattern recognition, trained to review the drone images automatically and flag anomalies consistent with a possible spill.
How it works
We started from a dataset of the client's own historical images, where experienced inspectors had labeled both real spills and common false positives: shadows, damp patches, terrain variations. That labeling work was key — a model is only as good as the examples it sees, and here the experienced inspector's knowledge was transferred, image by image, to the model.
On that foundation we trained a detection architecture based on convolutional neural networks, tuned to distinguish the visual signatures of a hydrocarbon over different types of soil from the harmless anomalies that used to require a second human look.
The system doesn't replace the inspector's final decision: it prioritizes it. Instead of reviewing 2,000 neutral images looking for an exception, the team gets a short list of flagged areas, ordered by probability, to review first. We didn't build something that decides on its own and asks the human to trust it blindly — we built something that filters out the noise and gives the inspector back their time and attention for the cases that truly need it.
What changed
The interesting part wasn't just that they gained speed, though they did. Monitoring was able to grow without adding people to watch screens, decisions moved earlier, and the team went from putting out fires to seeing them coming.
What those percentages hide is a change in stance. An operation that reacts moves after the problem. An operation that anticipates moves before it. In a business where a spill is measured in millions, that head start makes all the difference.
A note from today
This project we did a few years ago could today rely on far more mature tools. Computer vision models train faster and with less labeled data, inference can even run on board the drone itself in flight, and what used to require a hand-built data pipeline is now assembled on top of components that already exist. If we had to tackle the same problem now, the technical challenge would be smaller.
But the underlying lesson hasn't changed: technology is rarely the limit. The limit usually lies in how an operation processes what's already in front of it. What's still hanging in the air, besides the drones, is that question: how many problems that blow up today could be stopped if someone, or something, saw them in time?