If you live in Dallas, you probably already know allergy season hits hard. The city ranks among the top 20 allergy capitals in the United States, landing fourth for pollen score, over-the-counter allergy medication use and number of allergy specialists, according to a 2024 report from the Allergy and Asthma Foundation of America.
Now, researchers at the University of Texas at Arlington are using artificial intelligence to take the guesswork out of pollen tracking, starting with tree pollen.
In a recent study published in the journal Frontiers in Big Data, researchers at the university, along with scientists from Virginia Tech and the University of Nevada, Reno, developed an artificial intelligence system capable of distinguishing between pollen from fir, spruce and pine — three morphologically similar conifers — with about 99% accuracy.
While this system is a good step toward finding quicker ways to identify pollen, there is more work to be done before it’s ready for real-world use, said Mark Bush, a professor of biological sciences at the Florida Institute of Technology, who was not involved in the study.
“I think [the researchers are] doing everything right,” Bush said. “It’s just still such a big step,” he added, noting in the real world “there are thousands of pollen grains, and you’ve got all the crud that you find in a sample — all the dust and organic debris and everything else which might be obscuring the pollen grain.”
Assistant professor of biology at UT Arlington Behnaz Balmaki and collaborators at Virginia Tech and the University of Nevada, Reno, developed an artificial intelligence system capable of distinguishing between fir, spruce and pine tree pollen.(The University of Texas at Arlington)Pollen identification currently relies on expertise and visual inspection under a high-resolution microscope — a time-consuming and sometimes error-prone process, according to Behnaz Balmaki, an assistant professor of biology at UT-Arlington and co-author of the study. Finding a way to expedite this process was what motivated Balmaki and her colleagues to turn to artificial intelligence.
Teaching an artificial intelligence how to identify pollen grains required feeding it hundreds of images of these particles. Balmaki and her colleagues used pictures of six different types of fir, spruce and pine pollen grains sourced from the University of Nevada, Reno’s Museum of Natural History. This effort involved manually collecting pollen grains from the museum’s herbarium, placing hundreds of pollen grains — as many as 400 — on a slide and then snapping a photo under the microscope.
“We concentrated on conifers like fir, spruce and pine because they are very similar in shape and [we] wanted to see how the model works at identifying [them],” said Balmaki.
Masoud Rostami, an assistant professor at UT-Arlington’s Division of Data Science and Balmaki’s co-author, explained the artificial intelligence system learns to differentiate pollen types through multiple processing layers, similar to how these systems are trained to recognize faces by detecting simple features like shapes and edges, then progressively more complex patterns. This layered approach enables the artificial intelligence to catch subtle differences in fir, spruce and pine pollen.
Each layer, Rostami said, helps identify a different pattern in the grain. The more layers, the easier identification becomes, though the process does become slower and computationally more expensive over time.
Masoud Rostami is an assistant professor at UT Arlington s Division of Data Science and the study's co-author.(The University of Texas at Arlington)Balmaki, Rostami and their colleagues evaluated nine different artificial intelligence models and decided on one called called ResNet101, which they found offered a good balance of performance and computational efficiency, Rostami said. ResNet101 was 99% accurate in differentiating between the six kinds of fir, spruce and pine pollen using 101 layers. Some of the other models used as few as 18 layers, whereas others used as many as 201 layers.
Balmaki hopes this research will accelerate the pollen identification process and deliver precise, species‑level results. She and Rostami envision artificial intelligence systems like theirs will one day have applications spanning agriculture, public health, urban planning and beyond — anywhere pollen identification is important.
“Grass is a huge problem in Texas, especially for allergies,” Rostami said. “There is no paper, no publication, no research on [artificial intelligence systems] for grass pollen. That can be a very hot topic, focusing on the identification of grass pollen grains.”
Bush from the Florida Institute of Technology cautions that while the artificial intelligence system shows promise in pollen identification, scaling the model beyond the six kinds of pollen will be a monumental task. For regions like South America, where hundreds of distinct pollen types occur, teaching a system to differentiate between them requires a lot of data. Complicating matters are other organic particles, such as dust, that closely resemble pollen.
Balmaki and Rostami acknowledge there’s much work to be done to train their artificial intelligence to recognize more than six types of tree pollen. Their next step is to train it to recognize pollen native to North Texas.
“In the future, we plan to cover all of the United States, because later, when this data is ready, anyone can use it for any purposes they have,” Balmaki said. “For archaeology research, for allergy research, even for health, pollination, honey quality and forensic research, which is all very cool.”
Miriam Fauzia is a science reporting fellow at The Dallas Morning News. Her fellowship is supported by the University of Texas at Dallas. The News makes all editorial decisions.
English (US) ·