Newsroom

CORVALLIS - A team of experts at Oregon State University hopes to bring the powerful tools of artificial intelligence to the world of ecology and environmental protection, with a new $1.7 million grant from the National Science Foundation to create technology that can identify insects.

This application of "machine learning" could expand the frontiers of a computer science field known as pattern recognition, such as the systems that now use computers to identify fingerprints.

Fingerprints are easy, however, compared to identifying insects - flexible, three-dimensional objects that come in many shapes, sizes, colors and configurations. And there are a lot of them. In Oregon, a square meter of soil can contain from 100 to 300,000 individuals representing two to 200 species. But if the challenge is huge, so is the potential environmental payoff.

"When we perfect a low-cost, efficient method to use computers to monitor insect populations, it will revolutionize water quality monitoring, which will be one of the first applications," said Tom Dietterich, an OSU professor of computer science. "And in forest management, ecologists are hampered by lack of a way to easily measure insect populations and biodiversity. We might be able to speed up that process about 1,000 times."

What is needed is a system that can collect, manipulate, photograph and identify small insects, very quickly, accurately and in large numbers. And a long-term goal, the researchers say, is to create machine and computer systems that can be "retrained" for application to different sets of insects or a broad range of other pattern recognition problems.

One of the first insects the scientists will try to identify is stonefly larvae, which are known to be a sensitive indicator of stream health and water quality.

Changes in water quality over time can be tracked by monitoring changes in the composition of aquatic insect communities, providing more important information than a check on water at any one point in time.

"These larval stoneflies are sensitive to reductions in water quality caused by thermal pollution, eutrophication, sedimentation, and chemical pollution," Dietterich said. "They can provide us the first indication of a problem."

Researchers have known for some time the value of such monitoring, but the cost of collecting, counting and identifying the insects with highly-trained and often scarce experts is prohibitive.

A computer system that could do so would be of enormous value, even if only for this one application. And the broad fields of stream and terrestrial ecology, agriculture, forestry and many other areas could ultimately provide a myriad of uses for technology that could identify small, irregular objects quickly and accurately.

To expand the field of pattern recognition, scientists hope for advances in two areas - the use of "feature" dictionaries in which a computer can find, identify and use thousands of visual features from an object to determine its identity; and "relational appearance" methods, in which computers figure out what objects might look like under a wide variety of viewing angles and lighting conditions. The concept of "machine learning" or artificial intelligence implies that the computer systems, by themselves, will help decide which features and which views can be of the most help in making accurate identifications.

"These are difficult problems," Dietterich said. "If you look at the wing of an insect, for instance, there might be a number of lines between different cells, some of which are shared by several species and others that are specific to a single species. One reason this type of identification is rarely used for biological monitoring is that there are very few entomologists who even have the knowledge to do it." There may be fewer than a dozen taxonomic specialists in all of North America with the expertise to perform all of these analyses, the researchers said.

OSU is in an ideal position for this type of interdisciplinary research, the scientists say, because it can combine the skills of experts in a variety of computer science, natural resource and engineering fields. These types of research programs that cut across traditional academic boundaries are a growing trend at the university, Dietterich said. They will employ the efforts of a large number of undergraduate and graduate students, and the project directors also hope to integrate some of the experiments used in the work to train middle school and high school teachers in ecological field work.

Commercial spinoffs and new start-up companies using the technology that emerges from this research are also likely, he said.