K-State research takes flight to combat woody encroachment with precision mapping

Wildcats utilize machine learning to support grassland conservation.

By Emily Boragine
K‑State News and Communications Services

Hikers on the Konza Prairie may look up this summer and see something unusual in the air. Is it a bird? Is it a plane? No — it’s K-State research in action.

K-State researchers are employing aerial data in their mission to understand and manage the rapid spread of woody plants across the Great Plains.

Known as woody encroachment, the transformation of open grasslands into shrub and tree-dominated landscapes is impacting biodiversity, livestock forage, water resources and even wildfire risk.

“Woody encroachment is a pattern that is happening in grasslands all around the world, where areas that used to be grasses, wildflowers and other herbaceous species are seeing a rapid and large increase in shrubs and trees,” said Zak Ratajczak, assistant professor of biology at K-State. “That’s not to say that all of these changes are negative, but in a lot of these grassy ecosystems, shrubs and trees move in and change things drastically.”

Remote sensing meets machine learning

Published in the open-access journal Remote Sensing, a recent study led by K-State master’s student Brynn Noble and Ratajczak offers a cost-effective, open-access approach to detecting woody encroachment across landscapes as small as six-by-six feet. The system combines aerial imagery from federal programs with ground-based data collected by K-State researchers.

“We’re at the point where we can say if that area is a grassland, if it’s a shrub, if it’s a tree or if it’s an evergreen tree with around 97% accuracy,” Ratajczak said. “Meaning that 97% of the time we say that pixel is this type of vegetation, we’re getting it right.”

The accuracy is made possible through a unique combination of resources: access to consistent, high-resolution aerial data through the USDA and National Science Foundation; the computational power of open-source machine learning; and large training datasets collected through fieldwork at the Konza Prairie Biological Station — much of it done by undergraduate and graduate students.

The high-resolution aerial data are collected using LiDAR and multispectral imagery from aircraft equipped with advanced sensors. These sensors provide detailed, three-dimensional views of the landscape, capturing both the structure and composition of vegetation to monitor changes like woody encroachment with exceptional precision.

Interactive content by Flourish

Ratajczak said most people are familiar with how these machine learning models work, even if they don’t realize it.

“If you’ve ever been presented with a series of photos and asked to click all the tiles that show a crosswalk or bus, then you have participated in machine learning of images,” Ratajczak said. “Computers can learn to find important features in on-the-ground or aerial photos, but to do so accurately, they need to have lots of training samples to learn the types of patterns to pick. Just as importantly, they need samples to teach them what features to disregard. For instance, in early versions of this project the computer was thinking that the deep shadows behind trees were miniature evergreen trees, when they were really just shadows.”

To do this accurately takes thousands of samples, which is where K-State students and the Konza Prairie come in.

A training ground for technology — and students

The researchers developed the classification models using data from the Konza Prairie, a long-term ecological research site jointly managed by K-State and The Nature Conservancy.

“In this relatively small area of about 10 square miles, we have a lot of the major vegetation types that you find across eastern Kansas,” Ratajczak said. “And when we’re trying to train a computer to find a diverse array of vegetation types, that’s an invaluable resource.”

That means researchers — including students — can quickly validate or correct model predictions based on firsthand knowledge of the landscape. The students gather training data for the models and get experience working with a Geographic Information System, or GIS, and computer coding along the way.

One of those students is Brynn Noble, who co-led the research study. Her work alongside Ratajczak has made a major contribution to what the team hopes will become a regional, if not statewide, mapping tool.

“One of the catalysts for applying to expand this work beyond our site in Kansas is that federally and state-funded aerial campaigns are gathering this data, doing the really hard steps, and then making it free to the public, including us. We wouldn’t have been able to do this without that data from NSF and the USDA,” Ratajczak said.

Impacts beyond the prairie

The implications go far beyond classification.

“Now that we have these accurate measures of shrubs taking over grassy areas, we can estimate how many cattle or bison an area can realistically support in the long term,” Ratajczak said. “These data points have fed into our calculations of carrying capacity for our bison herd, and we adjust it accordingly.”

Other teams at K-State have used the same woody vegetation maps to assess fire risk, habitat selection by birds and small mammals, and even tick-borne disease exposure.

“This has been an unexpected turn for us. We’re on a lot of manuscripts about stream flow and bird diversity,” Ratajczak said. “We were part of a study led by Katy Silber and Alice Boyle, showing that when a small amount of woody cover moves in — which we’re able to detect with the data from these airplanes — the grasshopper sparrows’ use of the habitat starts to drop off. They just don’t really live there for the most part anymore.”

This occurrence is not just unique to bird species. Woody encroachment impacts a variety of habitats and will ultimately affect even more species.

An eye toward the future

Ratajczak said he hopes to share the resources through a website or an app for regional early-warning systems.

Eventually, the team envisions partnerships with landowners to gather additional training data and refine their models.

“If someone finds that we’re getting it wrong, we want to know that so that we can update our models,” Ratajczak said. “Our long-term goal is to start using satellite data to cover a larger area and project what we do into the past and the future.”

For now, this technology — built from open-source tools and federally funded data — is already making an impact in grassland conservation.

“This is real-world, transferable skill development,” Ratajczak said. “And it’s helping to increase our understanding of one of the major environmental challenges in grasslands.”