Predicting Winterkill: going from in-green sensors to the causes of damage

Winter damage is one of the most difficult things to manage on a golf course because, by the time you can see it, much of the evidence for what caused it is gone. 

You walk by a green in April and find dead turf. Was it crusted snow and ice that formed during a February thaw and refreeze? Was it saturated soils in a low spot of the green? Or was it exposure to wind where snow blew clear in January? Each is a different winter, and while we can currently make educated guesses about what caused the damage, we've never had the data to pinpoint what exactly happened.

It's this problem - teasing apart the causes of winter injury - that the WinterTurf project is solving.

Data for Predicting Winter Damage

Over the past six years, you've partnered with us to collect data that we never had before: 327 golf courses, 23 winters of historical surveying, on-course sensors, global satellite data, and over 1,600 course-season observations of what the winter did and whether the turf survived (Figure 1).

Figure 1: Damage observations combining historical damage surveys and end-of-season surveys. 

What this data has allowed us to do is build first-of-a-kind machine learning models that can accurately identify the locations where winter damage happened and forecast where it might show up in the spring.

Helping Machines Learn Agronomy

The major challenge we've faced now that we have a robust dataset has been that machine learning models don't understand plants well. Typically, these models are built using purely "data-driven" approaches, where all of the data is fed directly into the model, and then the model figures out what does and doesn't matter. 

This didn't work for winter damage though because it's caused by both the winter conditions that a turf stand is exposed to and the hardiness of the turf itself.

To overcome this problem, we've built models that incorporate the best of machine learning with knowledge of how a plant responds to its environment and acclimates during the fall in preparation for winter. We also built into the models what we know about the mechanisms driving winter damage to explicitly account for freeze-thaw cycles, saturated soils, high light and low temperature, among others.

Models to Forecast Winter Damage

The results from all of this data and machine learning are encouraging. Tested against winters the model had never seen, it ranks damaged course-seasons above undamaged ones about 75% of the time.

We also asked whether all of this work with the in-green sensors was worth it, or could we have only used satellite data? And what we found was that the in-green sensors improved classification and forecast skill to over 81%, but because we had fewer observations, that skill was unstable across model configurations.

What is causing damage?

In addition to classifying courses and forecasting damage, we also built a model to infer the mechanisms causing it. Going back to the early work of James Beard out of Michigan State University, we've known for a long time that winter damage is caused by a mixture of abiotic and biotic factors. Our attribution models currently reflect this for abiotic factors alone (see Figure 2), and while preliminary, illustrates where future work is headed.

Figure 2. Inferred dominant mechanism of winter damage for A) North America and B) Europe. These are not able to be directly ground truthed with the datasets we currently have acquired and will be a focus of future research.

What comes next?

As scientists, we'll of course always say we need more data, but in this case, the models show that's true. We need more observations of winter damage at the same sites we put in-green sensors so that we can continue to increase our forecasting skill. Our current modeling work shows that in-green sensors are the way to get accurate in-winter forecasts that will be able to reliably recommend specific actions to mitigate damage before it happens.

Right now, a particular weakness is that we don't have the data to fully validate our attribution model (displayed for a course in Figure 3). We're further expanding our remote sensing pipelines to help with this and doing more analysis of the weekly surveys, but without more in-green direct measurement, we're limited in what we can do. 

Figure 3. Damage risk graph through a winter for a single course.

Sign Up

If you're interested in helping make these systems better, there are three ways to join.

First, you can sign up for our pilot that will send regular reports to you about your golf courses' winter risk: https://z.umn.edu/interest-survey

We will be providing regular forecasts sent to peoples' emails about your current winter injury risk and recommendations for what you can do. See an example in Figure 4.

Figure 4. Example of WinterTurf regular report that superintendents can sign up at this link: https://z.umn.edu/interest-survey.

Second, you can request to host a sensor and donate to help us maintain and expand our fleet. Our large USDA grant funding this work ends in August of 2026, so we're running on a shoestring budget while we write grants for additional federal funding. Each in-green sensor costs roughly $4,500 to manufacture along with staff time to maintain the fleet and generate reports. 

Every bit helps, and those interested can contact Eric Watkins at [email protected] about how to make a contribution.

Lastly, we'd like to thank the National Institute of Food and Agriculture, U.S. Department of Agriculture, Specialty Crop Research Initiative under award number 2021-51181-35861 and the Minnesota Golf Course Superintendents Association who have supported this work in the past. Without those past investments, we wouldn't be to this point and every winter we're one more closer to consistent forecasting of winter damage.

 Banner photo: "Frost on Alderley Edge Golf Course" by Colin Park, via Geograph / Wikimedia Commons, CC BY-SA 2.0 (cropped)