ASU team uses AI to detect wildfires

Three ASU students have found an effective way to detect wildfires quickly

An ASU team has invented a software program that analyzes wildfire data from a surveillance drone, which has the potential to change the way emergency officials respond to wildfires. 

The project, known as Prometheus, recently won 4th place in the national Microsoft Imagination Cup competition in May, giving them a spot to compete in the world-wide championship event next month. 

The ASU team consisted of 3 students – Murong He, Facundo Santiago and David Azcona. He and Santiago just recently received their master’s degrees in business analytics, while Azcona is a PhD student who spent this past school year at ASU on a Fulbright fellowship.

Facundo Santiago said that the team was surprised at the minimal number of projects out there trying to address the fire detection problem, but when they tried themselves, they understood why – it's not that easy.

“We learned that tough problems sometimes require creative solutions, which is why we approached the problem from a different – nontraditional – point of view,” Santiago wrote. “Artificial intelligence, and in particular deep learning, has the ability to excel at a task that involves dealing with an image, recognizing patterns and identifying characteristics on it that are hard to see at a glance.”

Murong He said they use drones because they can get an aerial view of the area, while removing the human physical element. 

“That's sometimes a lot of the big challenges of detecting fires is you need a person to find the fire on a plane or helicopter or on the ground,” He said. “So you can be safe in in your facility or whatever, while serving the area.”

Santiago wrote that the computer can learn all the complexity of the task, they just need to provide data to correctly guide the learning process. 

Santiago wrote that in the case of drones, the idea came into their minds naturally.

“(I) was already working with this technology so it was easier for us to extend the work from that point," Santiago wrote. "In addition to that, drones are a technology (that is) getting cheaper and cheaper, so it becomes accessible to everyone … Drones allow us to map out big areas and remote locations without getting physical persons there.”

He said that after talking with firefighters and conducting research, the vast majority of wildfires are reported by humans.

“It's a little bit of a tedious process and you know, what, if there is a wildfire that starts in an area where there any humans and nobody happened to stumble upon it,” He said. “(Wildfires) can spread very quickly, in a matter of just minutes (depending on) strong winds or low humidity, and so to be able to have some sort of automated process to detect them without active human intervention would be really beneficial.”

Brian Gerber, director of the Emergency Management and Homeland Security program in the College of Public Service and Community Solutions, wrote that one of the key characteristics of the wildfire hazard is its unpredictability. 

“Using technology to improve situational awareness is important to improving overall response — and is particularly helpful in improving communication with the public which will lead to more effective evacuation practices,” Gerber wrote in an email. 

He said that there are Red Flag Alerts in areas deemed to have the conditions that would cause a fire to spread very quickly, so she said that those are great opportunities to add surveillance and monitor the areas carefully just in case a fire does break out. 

Santiago wrote that if you want to classify an image into a certain category, such as fire or non-fire, using machine learning, people can train a simple neural network with images of ongoing fires and other images without them, with the goal of the computer learning to distinguish them.

Santiago wrote that the characteristics that are required to perform categorization are too small with respect to the full image. 

Santiago wrote that they have to use a technique called object detection, where the image is partitioned or broken into small regions with the objective of driving the attention of the neural network to that small parts. 

In order for the computer to correctly recognize the fire, the team would need millions of images, which is something they can't provide, so they use transfer learning.

Using transfer learning the team doesn't start from scratch, but from a neural network. They used a network called "AlexNet" that can recognize hundreds of objects, but not fire, and that was training with millions of images.

Santiago wrote that the degree of success at fire detection indicates how much the model has learned about the fire, but retaining all it learned about other objects. This technique gives their solution durability to avoid being tricked or confused with other objects in the picture.

Santiago wrote that that's what differentiates their solution from any others.

“This is a brand-new research area, and somehow we are all learning as we experiment with it,” Santiago wrote. “Probably the hardest task was how to pack all the technology stack we were using and deploy it as a service that can be consumed by the end users. In technology, the difference between something that works and some things that don't are the details. You need to take care of details if you want to get it working.”

Santiago wrote that Prometheus may have its debut in the US as soon as September thanks to their partnership with the Arizona Department of Forestry and Fire Management. They are helping them get their UAV in compliance with all the legal requirements to operate.

Santiago wrote that the government spends millions of dollars fighting wildfires.

“A lot of resources are put on fighting them, but not that much on preventing them,” Santiago wrote. “The US National Weather Services publishes something that is called Red Flag Alerts, areas that are under risk of developing a fire. However, no extra preventions instruments are deployed in such areas. We are providing an alternative. That's why we say ‘let's put AI to work’ in these situations.”


Reach the reporter at jlmyer10@asu.edu or follow @jessiemy94 on Twitter. 

Like The State Press on Facebook and follow @statepress on Twitter.


Get the best of State Press delivered straight to your inbox.