Sign In

Communications of the ACM

ACM News

Here Comes the Rain Again

View as: Print Mobile App Share:
Recent flooding in New York City due to Hurricane Ida.

Machine learning models could be useful to future-proof urban areas prone to floods.

Credit: Jaymee Sire/Reuters

Floods are the most common natural disaster caused by weather, according to data from the National Severe Storms Laboratory of the U.S. National Oceanographic and Atmospheric Administration, which says floods are expected to become even more frequent, due to climate change. Often caused by extreme rainfall, floods can threaten lives and cause extensive damage to homes and infrastructure.

Flood prediction systems, however, can help limit their impact. At present, a combination of different types of models are used to develop forecasts, which can incorporate physics-based simulations to predict water levels based on surface flow in rivers, for example. However, flood prediction system often are complex and require a lot of computer power and time to produce forecasts. Furthermore, measurements such as rainfall are often required from weather monitoring stations on the ground, but extensive networks of such stations typically have only been set up in developed countries.

"Not every country has access to those resources, so predictions can be somewhat ad hoc depending on where you are," says John Kimball, professor of systems ecology at the University of Montana in Missoula.

Machine learning, therefore, is of interest to help make predictions more accessible. One of the advantages to using machine learning is that the models can be trained with remote sensing data, such as satellite observations, which are available globally and often at no cost. Machine learning systems also take much less time to generate flood predictions. "The ability to make predictions very quickly is particularly what's needed with these near-real-time forecasts to get meaningful information to folks on the ground," says Kimball.

In recent work, Kimball and his colleagues developed a machine learning model to forecast floods using satellite data. They focused on an area of eastern Zimbabwe and a part of neighboring Mozambique where cyclone Idai triggered a major flooding event in March 2019. Their aim was to see whether satellite information could have helped to predict the flood, and therefore could be used for future forecasts. "In that region, a better early warning system is needed because it's flood-prone," says Kimball.

The team used a standard machine learning model called Classification and Regression Trees that was available within the Google Earth Engine platform. They trained it using data on their region of interest from satellite observations for several years prior to the flooding event, which was available in the platform. Data from the Soil Moisture Active Passive (SMAP) satellite provided information about surface soil moisture, while data from another satellite called Landsat revealed surface water conditions. Global rainfall data, which did not come from a satellite, was used as well, to calibrate and verify the model. The system also was validated using data from other satellites.

Kimball and his colleagues found their model accurately predicted the flooding caused by cyclone Idai 24 hours in advance; meaningful, but less-accurate, forecasts also could be obtained as far as three days in advance. "Ideally, a seasonal forecast would be wonderful, but in terms of saving lives, even a few hours' advance notice can be critical," says Kimball.     

The team is now working on expanding the forecasting model to other regions. They also are developing a Web-based platform that could be accessed via smartphone, so their model can be used by people on the ground, such as local agencies helping to relocate people before a potential flood. "Even the general public could access maps to see what areas are flooding," says Kimball.

Other data also could be integrated into their system to provide more detailed flood forecasts. While their current model is able to predict how water will spread in a region over time, for example, it does not indicate the depth of any flooding. "We're working with partners within the region to incorporate other kinds of information, for example stream flow simulations, to make a more comprehensive assessment of flood danger," says Kimball.

Although machine learning is promising for predicting floods, it also faces challenges. In a recent paper, Dennis Wagenaar, a research fellow at Nanyang Technological University in Singapore, and his colleagues delved into current opportunities for machine learning, as well as concerns. Organizations involved with flood risk management, for example, often are aware that machine learning is an emerging tool they should use, but they may not have in-depth knowledge about it, or its use. Wagenaar said the paper "was aimed at these kinds of people who want to know where it is useful and where it is not."

The increasing availability of public data should prove an advantage for machine learning methods. The resolution of available satellite images is constantly improving, for example, while views of streets, such as 360-degree panoramas provided by Google Streetview, cover increasingly more locations and could become a valuable source of information. Wagenaar thinks algorithms could use street imagery to estimate the elevation of homes and therefore potential flood impact, an aspect of predictions that is currently lacking. "We can predict a flood, but we (often) don't know whether the water will enter homes or not," says Wagenaar.

On the other hand, Wagenaar and his colleagues believe machine learning systems should not completely replace existing methods. Physics-based models, for example, use well-established formulas to capture certain aspects of flood predictions, such as how water flows, and therefore probably can't be improved on using machine learning. Wagenaar thinks the focus should be on components of flood forecasts that can't be described by existing formulas, such as the impact of water depth on people or buildings. "I see more of a role for machine learning in some parts of the chain (of models) that are fairly complex," he says.

The interpretability of machine learning models is a concern, too. If a machine learning system predicts high flood risk in a specific area, for example, it could have serious implications, such as reducing the value of homes, or forcing people to evacuate/relocate. However, predictions can be based on relationships identified by algorithms that are not always obvious. "Flood models have an impact on communities, and you want to be able to explain those decisions," says Wagenaar.

Machine learning models could be useful to "future-proof" urban areas prone to floods. In new work, Wagenaar and his team are planning to train machine learning models on historical data of city growth to extrapolate how a city might expand in 50 years' time, initially using Manila in the Philippines as a case study. If agricultural fields will be turned into roads or pavement, for example, less water would be absorbed into the ground during extreme rainfalls, which would affect how water spreads out over land. And if such an area becomes more populated, more people would be at risk in the event of a flood.

"You might want to invest much more money actually (in a flood-protection wall), if you take into account that in the future, it needs to protect many more people," says Wagenaar. "I think that's very important for long-term planning."


Sandrine Ceurstemont is a freelance science writer based in London, U.K.


No entries found

Sign In for Full Access
» Forgot Password? » Create an ACM Web Account