The Way of the Future: How AI Predicts Crop Performance with Dr. Ryan Benke

Podcast: Play in new window | Download

Subscribe on iTunes | Android | Stitcher | SoundCloud | Spotify | RSS feed

Resources:
USDA-ARS Wheat Health, Genetics, and Quality Research Unit
Deep Learning with R
WSU Variety Testing Data

Episode transcription:

[MUSIC]

Drew Lyon: Hello, welcome to the WSU Wheat Beat podcast. I’m your host, Drew Lyon, and I want to thank you for joining me as we explore the world of small grains production and research at Washington State University. In each episode, I speak with researchers from WSU and the USDA-ARS to provide you with insights into the latest research on wheat and barley production.

If you enjoy the WSU Wheat Beat podcast, do us a favor and subscribe on iTunes or your favorite podcast app and leave us a review so others can find the show too.

[MUSIC]

My guest today is Dr. Ryan Benke. Ryan is a USDA-ARS postdoctoral research biologist in the Wheat Health, Genetics, and Quality Research Unit stationed at WSU in Dr. Xianran Li’s lab. He received a B.S. from the University of Jamestown in North Dakota and a Ph.D. in biochemistry at Purdue University in Indiana, where he investigated the metabolism of spontaneous cell death maize mutants.

In his current position, he is using artificial intelligence and machine learning to characterize the impacts of weather on wheat performance and to build models that can predict future crop performance across Washington.

Hello, Ryan.

Dr. Ryan Benke: Hey, Drew. Thanks for having me on the show.

Drew Lyon: So, I’m kind of interested in this. How do you transition from a Ph.D. in biochemistry to working with artificial intelligence?

Dr. Ryan Benke: Yeah, so, when I started, I had really no experience at all. And we were kind of thinking of ways that we could take our projects to new levels and we just kind of started working with machine learning and neural networks and things like that. And when I started, I was really struggling to pick up the concepts and apply it, but using books, online resources, like online artificial intelligence, I was able to–you know, talking to Chat GPT–kind of piece together how to work through this and how we could apply it to our research.

Drew Lyon: Okay. So, can you kind of define some of these terms? AI? Machine learning? Are they different, are they the same thing?

Dr. Ryan Benke: Yeah, so, I actually looked up the definition of this before I came on here because I was curious myself, because people kind of use the term to describe any kind of computational thing that is working. So, an artificial intelligence is something that–a computer that can apply human-type reasoning, basically finding patterns, learning, and then applying that learning to something else.

And so, with us in our machine learning where we’re using deep neural networks. We’re actually using neural pathways where multiple layers actually connect an input that you’re interested in–for us it’s weather–and then a desired output, such as crop yields and things like that. So, it’s able to recognize patterns, and with those patterns we’re able to apply those to data sets that it hasn’t seen for our prediction modeling.

Drew Lyon: Okay. So, you’re able to use large data sets and try to tease out, correlate–what we normally would do a correlation or something with?

Dr. Ryan Benke: Yeah, it really allows us to use multiple parameters. When I’m building just a model before I was using neural networks, I was kind of limited in the number of parameters I could consider because the models become too complex, the computational power is so much. But with my neural network, I’m able to take 100,000 variables and add it to the model to do these kind of predictions.

Drew Lyon: Okay, so, now describe to me what a neural network is.

Dr. Ryan Benke: Yeah. So, a neural network is–you can just kind of think of it as a set of steps. So, you have your input layer, which for us is our temperature or weather variables, and then in the subsequent layers, it starts to form connections, trying to predict your desired output. So, for us, our desired output is our crop performance. And so, through all of these subsequent layers, they form these–it’s kind of like a spider web where each connection is connected to something else, and each parameter is assigned a weight that can be used to predict your future parameter.

And so, when we have multiple–like all these like 100,000 temperature or environment parameters that we input here, each one is assigned a weight in each of these subsequent layers. And so, when I’m doing predictions later, I can apply an environment that I haven’t seen before. And based on the weights that [were] assigned earlier, it can kind of identify or make a prediction of what you would expect based on this new environment.

Drew Lyon: Okay. So, you’re using variety testing trial data, I understand, to try and make these connections with weather. Is that the idea?

Dr. Ryan Benke: Yes. So, all of the data that I’m using is open-source data from the WSU Wheat and Small Grains variety testing data. So, it’s really nice that I didn’t have to do any of these experiments [myself]. I had data to work with right away. And so, we have data dating back from 2002 to 2023 that we’re using, and it encompasses about 300 environments where multiple varieties have been evaluated at each of these environments.

And then we’re getting temperature, precipitation, [and] humidity data from a NASA power project that encompasses a lot of this area in Washington. And so, basically our input variables are these temperature parameters that are at these locations, and then our output is these crop predictions, these performance predictions that we’re trying to make. And so, we can kind of train the model on environments, like half of the environments, and then predict it on the other environments and things like that.

So, all the training comes from this internal data set from the variety testing trials.

Drew Lyon: Okay. So, you use the existing data to try to train the model and then see how well it predicts other data. And then when you get to the point where it’s doing a good job of that, then you can start projecting forward. Is that the idea?

Dr. Ryan Benke: Yeah. So, our initial experiments are all kind of internal validation where I’m using the environments that are available to predict the other environments that are available. The ideal scenario would be to apply this to environments that it hasn’t seen or future environments even.

There’s a postdoc in our lab, Laura Tibbs-Cortes. She’s actually using future temperature models to actually predict out to 2090 for her switchgrass project. So, she kind of internally trains the model on data that we have available from previous years and then kind of looks at how the temperature and environment parameters are going to be projected to be in the future. And then you can use those in the model to predict performance.

Drew Lyon: So, what are the weather parameters you’re using? You know, temperature would be one. Water in a dryland situation, like much of eastern Washington, I assume is a pretty important one. Anything else?

Dr. Ryan Benke: Yeah, so, we’re using temperature, relative humidity, and precipitation. And then we have a couple other mathematical transformations of those parameters. For example, if you take the day length and if so, if you take day length times temperature or things like that, we can start to create these unique parameters that aren’t just part of those four.

And another thing that is unique to our research is we’re not just looking at like a season average, we’re actually looking at critical windows after planting. So, 10 to 20 days after planting, what is the relative humidity? And then what is it 15 to 25 days after planting? And that’s how we accumulate all of these temperature variables, because we have them in these very discrete day windows that we can use in the model. And we actually piece out that when you look at these smaller windows that we call environmental indices, that the prediction power is actually quite a bit higher than if you just use the season average where you maybe have a very wet environment early in the season [and] it might get drier later, but we can actually capture–like 2 to 3 weeks after–we can start to find out environmental windows that are very important for predicting yield.

Drew Lyon: Okay. So, you know, as a wheat plant develops, there’s certain critical stages. Are you able to–in your timing–do you have some kind of programming for what stage? I’m thinking like, you know, heading or flowering–heat at that time might do more damage than would do early or later. And are you able to make those connections to critical growth stages, I guess?

Dr. Ryan Benke: Yeah. We try to keep our windows that we’re considering biologically relevant. So, when we’re looking at a trait like heading–I think the average for the population is around 70 to 75 days after planting–so, it doesn’t really make sense for us to look at a 120 days after planting environment variable. So, we’ve kind of restricted our analyses to these kind of biological windows and we start to see most of the windows that are most critical for these traits actually precede the trait by quite a few days. So, the trait that’s most correlated or most influential on heading actually happens quite a bit earlier than heading actually happens, which is nice because it gives us predictive power before the trait actually does appear.

Drew Lyon: Okay. Is this work mostly with spring wheat or are you working with spring and winter wheat?

Dr. Ryan Benke: So, right now I’ve been building everything based on spring wheat. The nice thing about this is it can be easily adapted to any other crop, as long as we have data for it. We started to work a little bit with the winter wheat thing. The vernalization period was a little bit hard to work around because everything is standardized off of January 1^st so I was finding a bunch of environment variables that might have been influential right after planting, but then when we start to normalize to January 1^st, it was a little bit more to work through. So, that would be the next step, probably the next system that will apply this to.

Drew Lyon: Okay, so how does AI aid your model predictions?

Dr. Ryan Benke: Yeah. So, one thing that AI really helps me to do is I can include so many more parameters in the models than we’ve done before. So, this technique was really kind of developed using one single environment parameter–so, finding one single environment window that was most correlated with the trait that we are trying to look at. And then we had all these, you know, hundreds of thousands of other variables that weren’t really being used in our model. And so, the goal was really to incorporate them into the model. And if we tried to just use a mathematical sequencing or something like that, the computational power was too much–I couldn’t do it. There [were] too many variables to consider; it would make your head spin.

With the neural network. I can really just give it all 100,000 variables and through [these] kind of like connections that it makes, it’s able to pick out which variables are most important without me having to do prescreening for them. So, if weather parameters are not actually important for the overall trait, it assigns a weight score of those to essentially zero, so they get dropped from the model as it’s forming these connections. As opposed to ones that are highly correlated, [they] actually get higher weights in the model and so they’re more influential in the model performance.

Drew Lyon: Okay. So yeah, that would save you a lot of time, wouldn’t it, rather than going through all that? Okay. So, how do you see this research helping growers?

Dr. Ryan Benke: Yeah. So that’s another question that I have–coming from biochemistry where I was almost working at the bench, just working on cellular, and now I’m more in an agronomy focus. The goal is–I haven’t really been focused too much before on how am I going to apply this to growers. And the way that I see it is that this data needs to be made accessible to growers.

I think, myself included, a lot of people see neural networks or artificial intelligence and they’re like, “I don’t do that. I don’t know how that’s going to work. I know it’s important. Somebody is going to figure it out later and apply it for me” type of thing. But I think that we can actually make this data publicly available.

And what our research actually allows is if you’re evaluating a variety, you know, you grow it, you will see what the value is. But we can actually start to make prediction models of like how a different variety would have performed in that location without the variety actually having to be grown. So, I think in the future, as this gets more developed, it will be nice where any grower in any area can kind of evaluate a couple of more varieties that they didn’t plant and get a good idea of like, “should I be growing this variety or is there another one that might have performed better based on these predictions?”

And then in the future, as the temperature and environment–or temperature model predictions–become a little bit better, the thought is, you know, “we can’t know the weather in the future,” but as that technology starts to increase, we can actually incorporate those future temperature models into our predictions to kind of predict future years, which I think is where this is headed as we continue with it.

Drew Lyon: Well, I could see growers having a lot of fun with that. What other variety could I have been growing and would it have done better or not?

Dr. Ryan Benke: Yeah, we had a conversation with Dr. Alison Thompson and she gave us that idea that it would be interesting to see–or growers might be interested in seeing what other varieties would have performed like in their field. And I think as this develops, we can give somewhat of an idea of how that would have performed compared to the actual performance.

Drew Lyon: I know I’ve had Amber Hauvermale on a couple of times talking about low falling numbers, and I’m wondering whether something like this could be used to predict whether an area is going to possibly have that problem in a given year.

Dr. Ryan Benke: Yeah. So, Bosen Zhang, who was a postdoc in our lab–actually his last day was yesterday–he actually has been applying some of this machine learning with falling number. I think there’s a manuscript–it’s in review, I think, right now. And he kind of mapped falling number predictions across the state of Washington in this same kind of general area.

The limitation so far is that the number of evaluations where falling number has been measured in a variety testing trial data was pretty minimal. Like I have 300 environments for my training, which is actually quite small when people are applying these neural networks for facial recognition and things like that, they’re using millions of training sites. So, as we can get more sites and more of these evaluations, the stronger these models are actually going to perform.

Drew Lyon: So, if somebody wanted to learn a little bit about AI or machine learning is there a resource you could recommend for them to take a look at?

Dr. Ryan Benke: Yeah. So, the book that I learned this out of is called Deep Learning [with] R. The package is actually called Keras, k-e-r-a-s. They have a really nice online resource that you can run through an example of how to take any kind of input data set that can predict any variable. I think the example online is housing markets in Boston; based on location is usually a pretty common one, what the price you would expect there. So, there were some nice tutorials to walk through on that website that I was able to adapt our data to as I work through it.

Drew Lyon: Okay, we’ll make sure we put a link to that resource on the web in the show notes so people who might be interested in it can go and check it out.

Dr. Ryan Benke: Yeah, it’s really easy to adapt once you get started.

Drew Lyon: Okay, well, you know, I heard your seminar on this topic and I thought it’d be of real interest to our growers. I appreciate you coming in and sharing this. I think it goes a little beyond my comprehension, but you give me some hope that biochemists can learn AI and machine learning, something totally out of the box; so maybe I can learn some of this stuff myself if I do enough reading.

Dr. Ryan Benke: Absolutely.

Drew Lyon: I appreciate you sharing this information with our guests and I think we’ll all keep an eye on where this leads because AI and machine learning seems to be the way of the future. And I think there will be a lot of interest in it as we go forward.

Thanks, Ryan.

Dr. Ryan Benke: Yeah, thank you, Drew.

[MUSIC]

Thanks for joining us and listening to the WSU Wheat Beat podcast. If you like what you hear don’t forget to subscribe and leave a review on iTunes or your favorite podcast app. If you have questions or topics you’d like to hear on future episodes, please email me at drew.lyon — that’s lyon@wsu.edu — (drew.lyon@wsu.edu). You can find us online at smallgrains.wsu.edu and on Facebook and Twitter [X] @WSUSmallGrains. The WSU Wheat Beat podcast is a production of CAHNRS Communications and the College of Agricultural, Human, and Natural Resource Sciences at Washington State University.

I’m Drew Lyon, we’ll see you next time.

[MUSIC]

---

The views, thoughts, and opinions expressed by guests of this podcast are their own and does not imply Washington State University’s endorsement.