New computer program rapidly detects homes, buildings damaged by wildfires

A new computer program powered by artificial intelligence takes mere minutes to determine whether homes and buildings have been destroyed by wildfires once the smoke has cleared.

Developed by scientists at Cal Poly in San Luis Obispo and Stanford University, DamageMap evaluates the destruction by scrutinizing post-wildfire aerial and satellite images.

Residents of the Santa Cruz Mountains who were evacuated during the CZU Lightning Complex fire in the summer of 2020 say that having such a program would have relieved them of a lot of stress and anguish.

“I spent days thinking, ‘My house could be burning right now,’” said Lisa Smith Beasley, a Boulder Creek resident who was ordered to leave her home during the devastating event.

Beasley’s home survived the flames. But she and other Northern California residents often waited weeks to hear from authorities whether their homes had made it through wildfires.

Andrew Fricker, a spatial ecologist at Cal Poly who co-developed DamageMap, said that the program should drastically reduce that waiting period. Once fully developed, the program would be available for free to the public and emergency responders.

“There are so many people in California who are impacted by this every single fire season,” Fricker said. “And it’s only going to get worse.”

Fricker and his colleagues at Cal Poly and Stanford published their peer-reviewed work on the program in the November issue of the International Journal of Disaster Risk Reduction.

Computer programs that detect damage from natural disasters from aerial and satellite photos have been in development for a quarter of a century. But most of them require that before-and-after photos be taken with similar angles, lighting and photo quality, an imperfect system that requires a costly and continuously updated catalog of images.

To determine which structures have burned, DamageMap relies solely on post-wildfire images and a digital database showing the locations of homes and buildings.

Over the past four decades, the number of burned acres and homes throughout the West has grown substantially, fueled in part by climate change.

California’s deadliest and most destructive wildfire — the 2018 Camp fire in Butte County — inspired the creation of DamageMap.

The inferno severely damaged Fricker’s childhood home in Chico, where his parents were still living, but luckily the wildfire didn’t burn the house to the ground.

During the evacuation, Fricker struggled to find out if the house was still standing. “I was frantically trying to download any satellite images that I could get, trying to get information for myself and our neighbors,” he recalled.

Wanting to prevent others from experiencing the same distress, Fricker gathered aerial images of the Camp fire’s destruction and Cal Fire’s door-to-door structure damage assessments. With this data, he and a team of Cal Poly undergraduates created a rudimentary prototype of DamageMap.

He took the prototype to Google’s 2019 Geo for Good Summit, where he met Krishna Rao, a Stanford graduate student. At the event, the two built an improved version of the program. And in the years that followed, they continued collaborating and recruited more scientists to work on the project.

Last year’s CZU Lightning Complex fire in Santa Cruz and San Mateo counties was California’s ninth most destructive wildfire. The event scorched more than 80,000 acres and destroyed nearly 1,500 structures, 911 of them Santa Cruz County homes.

Despite the evacuation orders, many of Beasley’s neighbors stayed behind. If a program like DamageMap had been available to continually update evacuees on the status of their houses, she said, more people probably would have fled to safety.

“It was the not knowing that made it so bad because you couldn’t look forward at all,” said M’Liss Jarvis Bounds, another Boulder Creek evacuee. She waited three weeks to hear that her house had survived the flames.

DamageMap works by first building a database of pre-fire home and building locations using satellite images or aerial photos. Then it looks at post-fire photos and decides which structures are damaged based on characteristics such as crumbled or blackened roofs.

The application uses “machine learning,” a form of artificial intelligence, or AI, to identify burned buildings.

Typically, computer programmers feed tens of thousands of images into a program so that it learns to identify specific patterns. Facebook, for example, uses machine learning to recognize faces and suggest people to “tag” in photos.

In developing DamageMap, researchers fed nearly 50,000 images of both burned and intact structures into the program, including photos from the 2017 Tubbs fire in Santa Rosa, the 2017 SoCal fire in Los Angeles and the 2018 Woolsey fire in Los Angeles and Ventura counties. Afterward, the programmers tested how well DamageMap had learned what a fire-damaged structure looks like by showing the application another 18,000 images from the Camp fire and the 2018 Carr fire in Shasta and Trinity counties.

The program correctly identified charred structures in the second set of photos at least 92% of the time in about 18 minutes, according to the published paper. But it made mistakes when trees or other objects blocked buildings from view and when roofs blended in with the surroundings.

Although it isn’t meant to replace post-fire assessments performed by people, technology that can quickly and accurately evaluate damage appeals to emergency responders.

“As the technology and machine learning technology develops, we’ll certainly use it in the unfortunate event of another Camp fire or Tubbs fire, where it mows down a lot of structures at once,” said Will Brewer, a geographic information system analyst and developer at Cal Fire.

For now, Fricker and his team are improving the program by feeding DamageMap more data to learn from. The more post-wildfire images it sees, the better it becomes at identifying damage.

The developers say a lack of funding is holding the program back from being available for broader use. So far, an $18,000 grant from Cal Poly has  been the main source of funding, but Fricker estimates another $80,000 will be needed to get the application up and running for the public.

Fricker said he needs to pay dedicated computer programmers to continue training it, and he must find a suitable online platform to host the program, which could be costly.

“The code works, and we have a lot of data,” Fricker said. “If people were motivated to get this out to the public for the next fire season, it could be done.”