Anna Underhill has been working with VitisGen since February 2019, when she joined the USDA-ARS Grape Genetics Research Unit as a research technician. She made the move to Geneva from her home state of Minnesota, where she completed her M.S. in Applied Plant Sciences with the University of Minnesota Grape Breeding & Enology program. Undergraduate studies at Iowa State University took her slightly farther afield with a semester in New Zealand, where she learned more about kiwifruit than she ever could have imagined. Anna works primarily to improve neural network analysis of powdery mildew for high-throughput phenotyping systems. What got you interested in grape breeding and phenotyping?
I grew up in Minnesota, where the shadow of Honeycrisp looms large and the university’s next apple releases are highly anticipated. I loved plants and science as a kid, so when I was thirteen, I decided what I wanted to be: an apple breeder! However, after learning how many apple breeders there were in the U.S. – three – I figured I should expand my goals at least a little. I studied Agronomy and Horticulture at Iowa State University with a focus in plant breeding, where I learned about a wide variety of crops and production systems. Although I was willing to work with any number of plants in grad school, fruit was still first in my mind, so I jumped at a chance to work with Dr. Matthew Clark in the University of Minnesota grape breeding program. There, I focused on image-based phenotyping methods to quantify fruit cluster compactness and identify related QTL. While working on my M.S., I attended the 2018 VitisGen2 meeting and learned all about the interesting research happening within the group. I was excited at the chance to continue research in grape computer vision and joined the team in early 2019.
What is your role with the VitisGen2 project?
My role in VitisGen2 is to support the work of the powdery mildew team, particularly in the area of computer vision and neural network analysis. We generate a tremendous amount of image data in our work, so analyzing it and keeping it organized is a key part of what I do. During the spring and summer, I help set up experiments when we receive leaves from collaborators around the country, and image these experiments several times to evaluate disease progression and severity. I also help refine the neural networks we use to quantify powdery mildew on leaves, as the more we can reduce noise in our analyses, the more accurate our results will be. Outreach is something I really enjoy, so I’m often participating in events that showcase our work – and the USDA/Cornell AgriTech as a whole – to stakeholders and the community.
What are some major challenges faced by the industry/breeders, and how will your work address them?
Powdery mildew is the most economically important disease in grape, with growers often having to spray fungicides ten to fifteen times a season to control it. The potential impact of powdery mildew-resistant grapes has been estimated at $1 billion in the U.S. alone, with grapes that could require only one or two sprays a year. Providing options for disease-resistant grapes with excellent fruit quality would be advantageous to growers, wine producers, farm workers, and consumers, so I don’t think the potential positive impact can be overstated. Our work is continually speeding up the process by which new grape populations can be screened for powdery mildew resistance, which helps pinpoint locations within the grape genome that are responsible for this resistance. Ultimately, this knowledge will improve breeding efficiency and contribute to the goal of high-quality grape cultivars with durable disease resistance.
What is the most exciting thing you’ve learned or done since starting work with VitisGen2?
I’ve learned so much about neural networks and machine learning since starting with VitisGen2. I focused on image processing in my Master’s, but the introduction of automated evaluation systems has been a totally new experience. This new topic has helped me come up with new research questions, like: how do we best train our neural networks to get the most accurate results? How should we construct image sets to make them as representative of a population as possible? As someone who comes primarily from a genetics and breeding background, it’s been a very interesting journey thus far and I’m excited to keep learning more.
What’s a typical day like for you, including both work tasks and what you look forward to when you get home?
A typical day for me starts fairly early (5am!) with a trip to the gym. When I get to work, I spend a lot of time in my office fine-tuning neural network analysis results, training neural networks with new image training sets, writing applications for image quality control in MATLAB, and ensuring our experimental data is backed up and organized. I can also frequently be found in the “robot lab”, where I image leaf discs on our Blackbird machines to look at powdery mildew severity. Occasionally, I’ll do wet lab work supporting our genetic mapping efforts. After work, I can often be found biking around Canandaigua Lake, whipping up delicious vegetarian dinners, reading my newest cookbook, or working through the endless backlog of shows I’ve been meaning to watch.