A Nigerian flood-forecasting system mistook 41,382 blue roof coverings for floodwater, creating a phantom lake across 68.4 square kilometres (26.4 square miles) of Lagos, according to a scientific paper published on Thursday. Researchers say the error fed thousands of false advisories to a popular weather app and wrongly marked tens of thousands of homes as being at risk.
The study, published in the African Journal of Machine Perception, examined satellite images and app records from April 2025 to March 2026. Its authors say the system’s AI repeatedly labelled bright blue tarpaulins and painted roofs as standing water, even on dry days. The paper calls the resulting map “one of the largest urban false-flood signals documented in the region”.
How rooftops became a lake
The system, built by Lagos-based SoroSentry Systems, combines satellite images with rainfall forecasts to draw flood-risk zones. Researchers Renáta Brázdil-Beran and Sname Joshua compared its map with field checks in six districts, including Oshodi and Agege. They found that vivid roof coverings, common on homes and workshops, closely matched the blue-green colours the model had learned to associate with water.
“It was not seeing a flood. It was seeing a city dressed in blue,” said Brázdil-Beran, a machine-vision researcher at the Lagos Metropolitan Institute of Digital Systems. The paper says the model then expanded small rooftop misclassifications into connected risk zones, joining buildings across blocks into what looked like a vast sheet of water.
It was not seeing a flood. It was seeing a city dressed in blue.
Renáta Brázdil-Beran, machine-vision researcher
The error had a practical reach: researchers linked 6,420 app advisories to the false map, with alerts sent to users in neighbourhoods that the model marked as threatened. The study found no evidence that the messages prompted an evacuation. But residents interviewed by the authors said some delayed errands or moved shop stock indoors after seeing flood warnings on their phones.
Alerts, corrections and safeguards
SoroSentry’s product director, Kadri Männik, said the company had withdrawn the affected map layer and replaced the image classifier in August. “The alerts were generated by a pilot feature that should not have carried this level of confidence,” Männik said. The company says it is contacting users who received the advisories and will publish a correction log by 16 October.
Sara Neretljak, an official with the fictional Lagos Flood Information Desk, said the agency did not issue evacuation orders based on the app’s map. Her office checked the flagged districts against rain gauges and reports from field teams, she said, but the paper shows the company’s risk layer was also viewed by two local insurance brokers assessing property exposure.
The study’s authors warn that the failure could recur wherever satellite models are trained on image sets that under-represent local building materials. They recommend checking AI flood maps against ground observations and publishing the confidence level of each alert. SoroSentry says its revised system now requires a second data source before it can send a high-risk warning.
For Lagos residents, the paper offers an unnerving explanation for a flood that existed only on screens: the city’s roofs had become, to an algorithm, a lake. The researchers say they will test the replacement model through the coming rainy season, when real storms will make it harder to distinguish a computer’s mistake from a genuine emergency.
