In recent years, residents of Grand Haven may have noticed an uptick in drones flying over city parks. While they may appear routine, those flights are part of a broader effort to address a growing threat to Michigan’s forests.

Across the Midwest, biologists, government officials and experts with the Michigan Department of Natural Resources are grappling with invasive species such as oak wilt and hemlock woolly adelgid — diseases that can rapidly devastate forest ecosystems. 

Now, a Grand Valley faculty member and his team of undergraduate and graduate students are developing artificial intelligence tools to detect these threats earlier and more efficiently.

“It happens so fast,” said Rahat Rafiq, assistant professor in the College of Computing and the graduate director of the college’s Artificial Intelligence program.

“You have these oak trees that have been here for hundreds, literally hundreds of years, and suddenly they are dying. This is like an epidemic here in the eastern part of the United States.”

Rafiq was first approached in 2023 by Heidi Frei, forest health specialist with the Michigan Department of Natural Resources, and Lawrence Burns, psychology professor, to see if Rafiq’s lab in GVSU's Applied Computing Institute could develop a solution to detect the early signs of oak wilt. 

Burns and his son, Nathan, have been vocal advocates for raising awareness of the disease and protecting the historic oaks in their hometown of Grand Haven.

“By and large, the best oak wilt management is prevention because when oak wilt is found, it is really difficult to manage," Frei said. “It's a lot easier to have early detection, and that's where this project comes in with Rahat.”

At the state level, Rafiq said, early detection remains the biggest challenge. Michigan has more than 19 million acres of forestland — over half of the state’s total land area — making identification of infected trees a needle-in-a-haystack problem. 

“This is huge for us,” Frei said. “It becomes very difficult to observe on foot and identify signs of oak wilt, or any tree disease, especially early.

“We have to look for these symptoms throughout the year, and that can be very challenging.”

By the time crews locate an infected tree, it is often too late to intervene.

Oak wilt spreads quickly. Once infected, a tree typically dies within a month and can infect other oaks within a 100-meter radius, Rafiq said. Beyond the ecological loss, late detection also carries a steep financial cost.

“If it's too late and you can’t do anything, then you have to cut down those trees and that costs about $6,000 per tree,” Rafiq said. “Not to mention the expenses in getting a team to that location.” 

In response, Rafiq and his team developed an approach that combines drone imagery with AI. Drones can cover several acres in a single flight, capturing high-resolution images of the forest canopy. Those images are then analyzed by AI models trained to detect early signs of oak wilt and other invasive species, including hemlock woolly adelgid.

Each image is geotagged, allowing DNR crews to pinpoint affected areas and respond quickly. 

To further streamline the process, they launched edgeforestry.com, enabling users — from private landowners to forestry professionals — to upload aerial property photos for rapid analysis and identification of invasive species.

“From my view, this is really going to help make better management decisions,” Frei said. “We do surveys on the ground, but it’s time- and labor-intensive to do it throughout the season. 

“This will help us identify the true nature of some of the problems we have. We can see if the problem is more widespread and that changes the solutions we can use.”

As the project advanced, new hurdles emerged. 

Hiring drone operators can be expensive for some communities, and internet access is often unreliable in remote parts of the state, making it difficult to upload large image files for analysis.

To overcome those obstacles, Rafiq’s team turned to government satellite data. Satellites orbit the Earth every few days and can identify parcels of land showing early signs of stress. Drones can then be deployed to those targeted areas for closer inspection.

“It's a two-tier solution where the first tier tells you where to look, and the second tier, you fly the drone to know exactly what disease it is,” Rafiq said. 

Rafiq’s team also addressed connectivity issues by building an interface that uses edge computing. Instead of relying on cloud-based systems that require stable internet connections, edge computing processes data locally.

“If you’re a DNR ranger in the Upper Peninsula and there’s no internet, you can take the SD card from the drone, plug it into the device, and the AI will tell you exactly where to go. You never need the internet,” Rafiq said.

Rafiq said the resulting AI models are lightweight, cost-effective and highly accurate. 

“For hemlock woolly adelgid, we’re approaching 98% recall, or accuracy,” he said. “That’s a big step toward protecting Michigan’s forests before it’s too late.”