Unleashing the Power of AI: University of Illinois and IBM Join Forces to Anticipate Flash Floods

Extreme weather is becoming more common, making the ability to better predict major weather events an increasingly important part of climate adaptation. Better forecasting already saves lives, as providing advanced notice of the need to evacuate or take shelter can prevent the significant loss of life a disaster may have otherwise caused, and artificial intelligence (AI) could prove to be a game-changer in this regard. AI’s ability to use data to make predictions from the patterns it finds suggests that its forecasting ability could drastically improve with further development. 

The tech company IBM is among those bringing AI into the sphere of mitigating the impact of weather-related disasters, and together with NASA, it's hard at work on revolutionary foundational models. These models learn from a broad dataset to make using them for many different tasks quicker and easier, as opposed to task-specific models that are trained with data designed to teach them to do one job. This way, a dataset doesn’t need to be painstakingly created for each new task, because the AI can apply the information it’s learned from other situations to teach itself.

Among the projects its working on, IBM partnered with the University of Illinois to develop a foundational model capable of anticipating heavy rainfall and flash floods in the Appalachian Mountains. 

The problem of predicting flash floods

Rain doesn’t fall the same way everywhere. The result depends on geography. The same storm that may drench flat land with few negative consequences becomes much more intense in mountainous regions, said Ana Barros, professor and engineering department head at the University of Illinois. The terrain in the southern Appalachian Mountains essentially captures the storm, causing an extreme amount of rain to fall at once.

“That's what the flash flood is about,” she said. “You are having a very strong storm system and having a very heavy precipitation storm that comes through and stays in place for a little bit — for 15 minutes or so.” 

Precipitation is one of the most difficult types of weather to predict, Barros said. It becomes even more complex to precisely predict the location, time, and amount of rainfall in mountainous regions because of the way storms interact with the terrain. 

Existing weather prediction models can be improved by training them on the wealth of data that is now available to allow for more advanced warnings. This is where the AI model that IBM and the University of Illinois are developing comes in. 

At present, warnings are only available about six hours before a flash flood is expected, Barros said. Even then, it’s hard to quantify just how bad a flood may be. She’s hopeful that the improved model will predict flash flood risk 18 to 24 hours ahead of time. Increasing the warning time empowers people, authorities, and emergency managers to plan accordingly and avoid dangerous situations. 

“We will not be able to actually stop the [storm],” Barros said. “But what we can make sure of is that people are not there.” 

AI model in progress

The new foundational AI model is still in its infancy. The University of Illinois team, led by Barros, is part of the IBM Sustainability Accelerator — a social impact program that supports projects using tech to help vulnerable populations address environmental threats. 

“We kick off every project with what we call the IBM garage, which is a process for unpacking these needs, hearing from the community, doing early design work for the technical solution that would come up with the project, and writing a technical roadmap for its implementation,” said Michael Jacobs, sustainability and social innovation leader for corporate social responsibility at IBM.

It’s not just about improving the AI model’s capabilities, either. Once that process is completed, the technology will be scaled up so its insights can be shared with the affected community, Jacobs said. From there, community members will be taught to interpret those insights.

“There are human limitations to all of this,” he said. “Even if we build the perfectly trained or tuned model for this specific use case, the question ... needs to be answered: How are the insights actually implemented or leveraged to help people?”

It’s one thing to know a flood is coming, but another to know what to do about it, Jacobs said. That’s why engaging locals and teaching them to use the tool is important. 

Improved outcomes are on the horizon

“Last year we had almost 100 deaths in the southern Appalachians,” Barros said. That’s what the team is aiming to prevent in the future. While the Appalachian Mountains are the starting point for the AI model, there’s potential to expand what is learned there to other regions. 

Though this AI-enabled advanced warning is focused on flash floods, “There's another level of forecasting that can be done, which has to do with landslides, because they often happen together,” she said. “There's a lot of hazard predictability and hazard response planning that this kind of work can help.”

Improved flash flood warning and response are proven to save lives. By increasing the amount of time people have to prepare and leave the area, foundational AI models are integral to climate change adaptation.