How Big Data Is Helping Predict and Fight Wildfires - Dell Technologies

With climate change rapidly increasing the danger and scale of wildfires, the dry lands of California are particularly vulnerable. The state’s largest ever, the Thomas Fire, in December 2017, raged over 282,000 acres, destroyed a thousand homes, and threatened the lives of tens of thousands more.

Yet in the midst of this aggressive firestorm emerged a surprising new tool to help the Los Angeles Fire Department monitor the fire and predict where it would go—and what it would do—next.

WIFIRE is an integrated system for fire analysis. Using computational techniques including signal processing, visualization, modeling and data assimilation, the web-based platform merges satellite imagery and real-time data from cameras and sensors to assemble a picture of the fire, the conditions around it, and its trajectory.

If the tool is assessing wind conditions, for example, WIFIRE integrates current data about wind and also gleans from thousands of earlier wind events what it is likely to do next. WIFIRE then converts this information into imagery within minutes when new data is acquired.

During the Thomas Fire, the app, with roads and topography, informed by current and past weather, historical fires and much more, refreshed every 15 minutes. Those viewing WIFIRE’s system, including firefighting teams, research developers, and—in some cases, the public— could see where the fire was likely to spread just before it got there.

With distinct predictive capabilities, it’s little wonder the Los Angeles Fire Department came to embrace the tool. The dynamic map, which appears on the web in several different versions—one for the firefighters where they can run scenarios and one for the public (without the predictive capability)—was critical in tracking the Thomas Fire’s rapid progression.

Even its developers, a team of supercomputer programmers and fire experts whose project was funded by the National Science Foundation, were surprised by how much the public turned to WIFIRE to stay up-to-the-minute current during California’s fires this year. Firefighters used the app, but so did residents preparing for evacuation and worried families and friends both near and far. In the 39 days the fire raged, the app received about 10 million individual views.

It also appeared on network news, all a little stunning for a team that set out to build a predictive interface for firefighters and emergency planners.

“It just went with word of mouth – I mean, it’s a research project,” WIFIRE’s principal investigator, Ilkay Altintas said. “But the number of times the public accessed it – the scale of it being able to affect hundreds of thousands of people – that made me most proud.”

Altinta, the Chief Data Science Officer at the San Diego Supercomputer Center at the University of California San Diego, was awarded a grant in 2013 to develop an end-to-end infrastructure for real-time and data-driven simulation, prediction, and visualization of wildfire behavior. The WIFIRE teamincludes fire behavior and management experts from the University of Maryland, as well as several Supercomputer Center data and web technologists, government policy managers, private industry, and firefighters.

Beneath the WIFIRE platform that makes data-driven predictions and displays information within 15 minutes of the fire’s arrival is machine learning software that makes sense of, applies, and learns from the massive amount of data.

“A lot of our data is managed and prepared outside of our platform, which is why we were able to model down to minutes,” Altintas said. “Our group’s expertise is automated workflow so everything is automated and everything is saved so that we can always prove how our model was generated. It makes things a lot more accountable.”

And that underlying data is massive: It includes images gathered from U.S. Forest Service sensors, from the university’s HPWren project, and from weather stations established by San Diego Gas & Electric. It also includes weather data from the National Weather Service, geospatial and topography data, and figures about the behavior of earlier fires.

“We are using various analytical, machine learning, Geographic Information Systems (mapping) and signal processing techniques to curate and visualize the data sets,” Altintas explained. “Such techniques, we hope, are changing the way we use data for fire modeling and similar hazards modeling as fast as the real-time measurements flow in.”

But one data set that doesn’t exist, yet, in real time is information on the brush, trees, and other potential fuels for a wildfire. Altintas said that student projects are working to see if applying deep learning techniques to high resolution satellite imagery could translate into more accurate identification of fuels, which are vital to understanding a fire’s path.

“We have developed techniques to compare the models over time to the data we are receiving from the fire and we can dynamically adjust the model as we receive information,” she said. “That’s the real innovation here: dynamic adjustments over time.”

LAFD Commander Carlos Calvillo now calls WIFIRE an extremely valuable tool. But the project first had to prove its worth.

“We were disbelievers at first; we’ve seen other fire modeling programs,” he said. “My GIS expert is now one of the biggest believers, and he was very apprehensive about it when he first looked at it.”

The first time the department used WIFIRE, it ran parallel to the firefighting effort but wasn’t relied on. Afterward, firefighters reviewed what WIFIRE reported and predicted and found it quite accurate, Calvillo said. The firefighters eased it into their toolkit and over time, it won their trust. It soon went into full use, and firefighters are now trained on how to use the platform. The WIFIRE team worked hard to make it intuitive and easy to use, he said.

“What the tool gives us is within the initial action phase, really within minutes, we’re able to run a model of the incident and have a computer model of where this fire is predicted to spread,” he said. “For the initial action commanders who are trying to make decisions, when they’re going off their experience, they’re going off their gut, this can confirm or reconfirm how they thought the fire was going to spread or give them indications of where it will go that they haven’t even thought of.”

That information, which Calvillo says has been very accurate, helps incident commanders close to the fire makes decisions about where to place and how to use apparatus and firefighters, how to prioritize evacuations, and how many resources they will need.

“The logistical support (behind the visible firefighting effort) is unreal,” he says. “What people see is the boots-on-the-ground putting the wet stuff on the red stuff. But the planning effort takes tremendous resources.”

The app can also affirm the decision experienced commanders must make – from where aircraft should dump water to where firefighters and trucks should move to next.

“It’s these types of critical decisions that the incident commander is making that this tool can help validate,” he said. “Much of our task is keeping ahead of the fire, keeping it out of communities, evacuating people, and keeping the public informed.”

Firefighters, whose version of WIFIRE contains far more information than the public interface, are able to see current predictive models and run scenarios. The tool helps them answer questions like: If we stop the fire here, will it move somewhere else, and where? What if the wind shifts? Where can we put the most effective blocks? Which communities are in its path and where are the most effective places to try to prevent the fire’s movement?

“We’ve been able to get this information in a very compressed period of time which we’ve never been able to do before. It’s a very powerful tool,” Calvillo said.

When Altintas, the principal investigator, proposed the project to the National Science Foundation (NSF), it was well received – funded with $2.7 million in 2013 for five years. Irene Qualters, the office director for the foundation’s Office of Advanced Cyberinfrastructure, said the project appealed to the foundation because of its collaboration among firefighting and the academic disciplines of computing and data science. (Altintas had also already teamed with the U.S. Forest Service to use the federal agency’s sensors.)

“It had a very strong plan with equally strong outreach. The principal investigator came in with a number of others across disciplines working on an important problem,” Qualters said. “We knew from the beginning that [Altintas] had the right parties engaged.”

What’s more, the foundation prefers projects that give the taxpaying public access to the results, and the WIFIRE user numbers during the Thomas Fire proved the public was engaged.

“Dr. Altintas is a data scientist and she has built a platform that is scalable, meaning others can use the platform with their own data,” she said. “We’ve also seen that the models are getting better.”

Yet, one of the ironies of the project is that its “successes” are measured, in effect, by what didn’t happen—what didn’t burn. While WIFIRE can demonstrate how effective it was at helping firefighters stop the fire, it’s harder to measure what did not get destroyed.

“When firefighters are able to predict the fire’s path more effectively, they can protect communities and homes more effectively. Running fire models that leave out the fire suppression efforts after the fire can show the impact of firefighting that they were able to save so much,” Altintas said. “The success of those firefighting efforts is the disasters that didn’t happen, the people who didn’t die or lose their homes.”

With NSF funding coming to an end later this year, Altintas is looking for a business model to get the WIFIRE program to all the organizations and people who could use it, including fire departments, city, state, and federal land agencies, and joint task forces in California and every other state facing wildfire threats.

“In California, this type of infrastructure is very useful,” she said. “Information networks and predictive systems like this are really cheap to build compared to the losses from fires. What can sound like a lot of money initially can be minimal overall when you think of what you save through more informed decisions.”