A water leak can run beneath a paved road for some time before anyone sees a tell-tale puddle. Researchers are now asking whether satellite water leaks detection could help engineers find trouble earlier by looking for unusual signals from the ground above buried pipes. In a 2026 study, scientists combined radar imagery with machine learning to identify suspected leaks beneath paved surfaces. Their results suggest a useful way to narrow a search, but they also reveal a substantial problem with false alarms.
The research by Shiva Arabi and David Grau, published in the Journal of Environmental Management, used radar data to look for changes associated with underground moisture. The best-performing method reportedly identified 81 per cent of actual leaks while cutting the inspection effort in the studied scenario by 93 per cent. That is a promising screening result, not evidence that a satellite can see a pipe crack directly or identify every leak. The authors explicitly report many false positives, and their work concerns particular environments.
Why water leaks are not visible like potholes
Water networks hide much of their most important infrastructure beneath roads, pavements and buildings. A small problem can be difficult to identify because the visible surface may remain dry while water spreads through the material below. Traditional inspections can include listening for escaping water, measuring pressure, looking for changes in supply flow and investigating suspect areas on the ground. Each method offers clues, but a large network is expensive to check continuously.
A satellite does not photograph the inside of a buried pipe. Radar sensors record information about the surface and the way signals interact with the environment. Changes in moisture can affect the data, potentially leaving patterns associated with water escaping underground. Machine learning is used to search for unusual patterns that deserve investigation. It is an indirect measurement, so other causes of changing ground conditions may look similar.
This distinction matters for the public story. The prospect is not a camera looking through concrete and locating a broken joint as though it were visible behind glass. It is a system that may identify suspicious areas from above and help a maintenance team choose where to begin. The human confirmation remains essential, especially when a street excavation could disrupt residents, traffic and businesses.
What the researchers tested beneath paved surfaces
The study focused on radar data in arid settings with relatively sparse vegetation. That matters because an open, dry environment may reveal moisture anomalies differently from a city with frequent rain, tree cover or complicated underground drainage. The scientists explored a form of anomaly detection in which a model looks for unusual patterns rather than simply matching every image to a large library of confirmed leak photographs.
The reported detection of 81 per cent of actual leaks is encouraging because a screening system could potentially direct scarce inspection resources towards likely trouble. But the paper also reports a high number of false positives. A false positive is a location flagged as suspect even though it is not the leak being sought. Such warnings may still consume time and money, and a utility cannot excavate every unusual radar pattern without checking other evidence.
The reduction in inspection effort is therefore a finding about the research workflow, not an assurance that every water company would save the same percentage of labour or expenditure. Transferring a method to another network means testing whether the environment, pipe locations, ground material and data quality are comparable. Results from a particular study should be read as evidence of potential, not treated as a universal operating figure.
Why this is a public infrastructure story, not only a technical one
Water lost from the network still has to be treated, pumped or replaced. When leaks are not found quickly, they can contribute to wasted resources and damage around buried infrastructure. A company that knows where its network is most likely to be leaking can investigate more intelligently. That could support more targeted maintenance and reduce unnecessary searches, although the research does not prove it will immediately change any customer's water bill.
From a resident's perspective, the problem is familiar. A road is opened for maintenance, a leak is repaired and the surface is reinstated. Finding the right stretch of pipe before opening the road would make the operation more efficient. The satellite method's possible contribution is to improve the choice of where engineers look first. It does not replace pipe maps, measurements, trained staff or the physical repair itself.
Environmental conditions also affect how urgent a suspected leak may be. A visible burst flooding a road requires a different response from an uncertain pattern in remote radar data. Maintenance teams must combine observations with operational knowledge and prioritise risks to supply, property and safety. A computer model can support that judgement without becoming the sole decision-maker.
Why the difference between a warning and a repair matters
A machine learning model returns a pattern or probability, while the water network requires an action. The gap between those things is where engineering discipline becomes important. Operators may cross-check an alert against pressure records, unusual customer reports, past repairs or additional satellite observations. If those sources agree, the suspect area becomes more compelling. If they disagree, engineers may decide that the radar signal reflects something else.
One benefit of an automated screening method is consistency. It can apply a method across a large area without relying on a worker to visually examine every image individually. The downside is that it may repeat the same mistake systematically. A model sensitive to a particular environmental feature could keep highlighting harmless locations unless its performance is checked against actual repairs and confirmed leaks.
Good deployment would therefore track both the leaks found and the investigations that led nowhere. Reporting only the successes would make the tool appear more accurate than it really is. The false-alarm rate, the missed-leak rate and the practical cost of each inspection all matter when deciding whether to purchase or deploy such a system.
Water leaks at home are a separate, simpler problem
The US Environmental Protection Agency's water-saving advice explains ways households can find ordinary plumbing leaks using meter readings and checks on taps, toilets and fittings. That advice concerns buildings rather than buried municipal pipelines. It serves as a useful reminder that not every water problem needs satellite imagery or sophisticated AI. Sometimes the most effective detection method is a careful inspection of a familiar piece of plumbing.
A domestic leak detector, an unusual bill and a network-level radar study operate at very different scales. Combining them into a single claim that AI can spot every hidden leak would be misleading. The interesting research here is about large infrastructure where direct inspection is difficult and where better evidence about location could help maintenance teams.
AI is learning to look for changes that humans cannot easily see
Satellite monitoring already features in environmental work where changes over time matter. LiveAIWire has reported on AI flood forecasting and methods for recognising changes across the Earth's surface. Both involve interpreting observations with models, but their questions are different. Predicting a flood, mapping a changing landscape and flagging possible underground water loss should not be treated as interchangeable tasks.
The value of remote sensing often comes from combining measurements with other forms of evidence. A satellite can examine an extensive area repeatedly, whereas a local engineer can understand a network's valves, usage patterns and repair history. The two perspectives can support one another. The strongest system is likely to be one that brings these kinds of evidence together rather than elevating any single algorithm above the others.
Longer-term forecasting is another distinct task, as LiveAIWire's coverage of AI methods for predicting weather beyond the immediate forecast window has shown. In each field, useful predictions depend on clear limits and appropriate verification. A model that works on its chosen benchmark may need substantial testing before it provides reliable advice in another setting.
The next question is whether it works after the rain
The case for further research is strong because locating leaks is a costly, recurring task and buried pipes are inherently difficult to inspect. A practical trial would need to assess different climates, road surfaces, vegetation and pipe materials. It would also need to compare the cost of satellite analysis, follow-up inspections and actual repairs with the methods utilities already employ.
For now, the responsible message is that researchers have demonstrated a promising way to flag suspected leaks using radar data from space, while exposing significant limits. It would be an error to conclude that the technology has already solved buried water loss. But being able to shrink the search area before sending engineers into the street is a useful enough ambition on its own.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.