Spatial machine learning and explainable artificial intelligence predict urban road flood vulnerability in coastal West Africa

In the coastal cities of West Africa, the ground is shifting under the feet of those who build and maintain them. Rapid urban growth is paving over natural landscapes, turning soil that once drank rain into hard surfaces that push water away. At the same time, the climate is changing, bringing heavier storms that overwhelm old drainage systems. For the engineers tasked with keeping roads open and safe, this creates a dangerous puzzle. Traditional methods for predicting where a road might flood rely on detailed, expensive surveys of the ground and complex physics simulations. These approaches work well for a single bridge or a major highway, but they are too costly and slow to apply to the vast, tangled networks of smaller streets that connect entire neighborhoods. When a city cannot afford to map every inch of its road network, it often reacts only after the damage is done, repairing roads only after they have been washed away.

A new approach offers a different path, one that relies on the vast amounts of free data already available from satellites and open maps. Researchers have developed a method that combines these open data sources with a type of computer learning designed to be transparent. Instead of treating the computer as a mysterious "black box" that simply outputs a risk score, this method uses a technique called explainable artificial intelligence. This tool peels back the layers of the computer's decision-making to reveal the specific physical reasons why a road is vulnerable. It allows engineers to see not just that a road is at risk, but why it is at risk, turning abstract probabilities into concrete engineering instructions.

The study focused on the road networks of coastal West African cities, where the threat of flooding is acute. The researchers started by taking the digital map of the roads and breaking them down into small, manageable pieces, each exactly one hundred meters long. For every single piece of road, they gathered a set of facts from free, global databases. They looked at how high the ground was, how steep the slope was, how much of the surrounding area was covered in concrete or buildings, and how much rain had fallen in the past. They also checked historical records to see which road segments had actually flooded before. This created a massive dataset where every road segment was paired with the conditions that led to its failure or its survival.

To make sense of this data, the team trained a powerful computer model known as a gradient-boosted decision tree. Think of this model as a highly skilled analyst that can spot subtle patterns in a mountain of information that a human might miss. The model learned to predict whether a road segment would flood based on its physical characteristics. It performed with remarkable accuracy, correctly identifying vulnerable roads in the vast majority of cases. However, the true breakthrough was not just in the prediction, but in the explanation. By using the explainable artificial intelligence tool, the researchers could ask the model to show its work. They discovered that the computer was not guessing; it was relying on specific, measurable physical limits.

The findings overturned a common assumption in urban planning. For decades, city officials have often assumed that major, primary roads are the most critical to protect, while smaller, local streets are less important. The data showed this to be false. The computer model found that the administrative label of a road—whether it was a main artery or a small feeder street—had no real effect on whether it would flood. A small road in a low-lying, flat area with high amounts of concrete nearby was just as likely to fail as a major highway in the same spot. The factors that truly mattered were the local physical conditions: the height of the road above sea level, the steepness of the ground, and how much of the surrounding land was covered by buildings.

The study pinpointed exact thresholds where the risk of flooding jumps dramatically. Roads sitting at an elevation lower than twelve meters above sea level were found to be in severe danger. When the ground slope was flatter than two and a half degrees, water had nowhere to go, leading to dangerous pooling. Similarly, when more than sixty-five percent of the land within a short distance of the road was covered by buildings or pavement, the water had no place to soak in, forcing it to rush over the road surface. The model also identified that when a single day of rain exceeds one hundred and forty-five millimeters, the structural base of the road becomes saturated and fails. These are not vague warnings; they are precise, physical tipping points.

This work provides a clear blueprint for cities that lack the budget for expensive field surveys. Instead of waiting for a disaster to strike, municipal engineers can now use these specific numbers to guide their decisions. If a road is in a flat area with a slope of less than two and a half degrees, the city knows it needs to raise the roadbed or improve the drainage immediately. If a neighborhood has become so paved over that more than sixty-five percent of the surface is impermeable, the city knows it must install permeable surfaces or larger drainage pipes. The research suggests that by focusing on these physical realities rather than administrative road categories, cities can direct their limited funds to the places where they will do the most good.

The implications extend beyond just fixing roads. By understanding exactly where and why infrastructure fails, cities can plan for the future with greater confidence. The researchers propose that these findings should be written into local building codes and drainage design rules. When a new road is planned, or an old one is being repaired, the design should be based on these physical thresholds. If the ground is too flat or the surrounding area too paved, the engineering specifications must change to compensate. This shift from reactive repair to proactive, data-driven design offers a way to build resilience in some of the world's most vulnerable coastal areas.

Ultimately, the study demonstrates that high-quality climate adaptation does not always require expensive new technology or massive field teams. By combining freely available satellite data with smart, transparent computer models, it is possible to see the hidden risks in a city's infrastructure. The result is a practical, actionable guide that tells engineers exactly what to fix and how to fix it. In a world where climate change is making extreme weather more common, this ability to predict and prevent failure is a vital tool for keeping communities connected and safe.