Library

260: AI Finds New Grape Growing Regions as Climate Changes

In the face of climate uncertainty, growers wonder which grape varieties will flourish in their regions in the future, or if any will grow there at all. Joel Harms, a Ph.D. student in the Department of Bioresource Engineering at McGill University in Australia is using artificial intelligence to simulate the potential to grow pinot noir in different regions of the world that are currently considered too cool. The project mapped 1,300 varieties to 16 different points of climate data including temperature, precipitation, and growing degree days. The findings could play a crucial role in identifying the winegrowing regions of tomorrow.

Resources:        

Vineyard Team Programs:

Get More

Subscribe wherever you listen so you never miss an episode on the latest science and research with the Sustainable Winegrowing Podcast. Since 1994, Vineyard Team has been your resource for workshops and field demonstrations, research, and events dedicated to the stewardship of our natural resources.

Learn more at www.vineyardteam.org.  

Transcript

[00:00:00] Beth Vukmanic: In the face of climate uncertainty, growers wonder which grape varieties will flourish in their regions in the future, or if any, will grow there at all.

[00:00:13] Welcome to Sustainable Wine Growing with the Vineyard Team, where we bring you the latest in science and research for the wine industry. I'm Beth Vukmanic, Executive Director.

[00:00:23] In today's podcast, Craig McMillan, Critical Resource Manager at Niner Wine Estates, with longtime SIP certified vineyard and the first ever SIP certified winery. Speaks with Joel Harms, PhD student in the Department of Bioresource Engineering at McGill University in Australia.

[00:00:42] Joel is using artificial intelligence to simulate the potential to grow Pinot Noir in different regions of the world that are currently considered too cool.

[00:00:52] The project mapped 1, 300 varieties to 16 different points of climate data. including temperature, precipitation, and growing degree days. The findings could play a critical role in identifying the wine growing regions of tomorrow.

[00:01:07] Want to be more connected with the viticulture industry but don't know where to start? Become a member of the Vineyard Team. Get access to the latest science based practices, experts, growers, and wine industry tools through both infield and online education so that you can grow your business. Visit vineyardteam. org and choose grower or business to join the community today. Now let's listen in.

[00:01:34] Craig Macmillan: Our guest today is Joel Harms. He's a PhD student in the Department of Bioresources Engineering at McGill University. And today we're going to talk about mapping global future potential for Pinot Noir cultivation under climate uncertainty using generative AI.

[00:01:51] Bye. Bye. This is a really interesting topic. I came across an abstract from a recent ASEV meeting and I was like, I just have to know more about this. This just sounds too interesting. But welcome to the podcast, Joel.

[00:02:04] Joel Harms: Okay. Thank you very much. Thank you for having me.

[00:02:06] Craig Macmillan: What got you interested in this topic in terms of this wine grape region? Stuff.

[00:02:12] Joel Harms: I think it was more about I wanted to build models that are useful, I guess, broadly useful in vineyard management and like establishing new vineyards and like kind of covering some of the base problems. Initially, my thought was, how can we. see which grape varieties are alike.

[00:02:32] How can we like make a representation of them in like a latent space. But then I found out , if I do that, that's, you know, somewhat useful, but if I take that just a step further, I could just connect it with climate data already. And then we would have a model that could, be used for prediction and it would be so I guess. How do I say like broad or general enough so that you could apply it in any environment. So like any climate can be used to predict any grape suitability matrix, which is quite nice. And so then I thought, no, let's do it. Let's try that.

[00:03:11] Craig Macmillan: So your colleagues and yourself did some simulations, as we just mentioned specifically around Pinot Noir and the potential to grow it in different parts of the world that currently are considered too cool. Tell us exactly how you went about this.

[00:03:25] Joel Harms: The abstract is kind of a case study on one application of, These models that we built. So we built very general grape variety recommender systems based on climate. And so we wanted to show a cool application globally. This can be applied to find regions that will be too hot in the future.

[00:03:43] So we built the AI models first starting from looking at where grapes are grown and tying that together with what climate is there regionally. Unfortunately, you know, we can't use like very precise climate data because we don't have the exact location of each grape variety in each region.

[00:04:02] Craig Macmillan: hmm. Yep.

[00:04:03] Joel Harms: Yeah. So therefore, we use larger climate data. So like at 50 kilometer resolution, which is still helpful to, I think, gather overall trends, not so much, you know, to plan an individual vineyard probably, but just to see like in which areas maybe there would be. in the future interesting vineyard sites.

[00:04:23] Just like kind of as like a pre guidance sort of model. And then we, tested it. We tried to validate this model and then we presented a first case study with Pinot Noir because we were presenting in Oregon at the ASEV conference. So I figured, you know, might as well do Pinot Noir if we're already in Oregon.

[00:04:43] Craig Macmillan: Can you explain to me the artificial intelligence piece of this? I mean, you hear about it and you know, kind of what different types of AI do. I don't think a lot of people realize that, you know, that's a very general concept and people have designed particular tools for particular reasons.

[00:05:01] So, in this case, what exactly was the AI component? What's inside the box, basically? How does it work?

[00:05:07] Joel Harms: First off, I guess to explain for listeners , cause AI does get thrown around a lot and it's hard to know what that actually means. So when we're talking about AI, it's usually we're tying some sort of input data to some sort of output data. And we're teaching a very complicated mathematical function to map one to the other.

[00:05:25] So like kind of a correlation. But it's not a simple correlation. That's why we need these models and that's why they're pretty fancy.

[00:05:31] So in our case, we're using an AI that was inspired from the community of medical science, where similar models were used to connect, for example, the ECG measurements of a heart with like scans of the heart.

[00:05:50] And then Trying to tie both of those datas together and to reconstruct them again to see if, like, you could find correlations between those and maybe if one of them is missing, you could, , predict what it would look like. And so, since this is a very similar problem, , and we have similar input data in the sense of, we have grapes, which grapes are grown where, and we have what is the climate there, roughly.

[00:06:13] So we can tie that together and try to connect both of those types of data and then get an output of both of those types of data so that we can go from grapes to climate and climate to grapes in the same model. So we have these , you could say like four models. that are tied together at the center. So input grapes, input climate, then in the center where they get tied together and then output grapes, output climate. And so we train it to, reconstruct it from this combined space where we like, Scrunch it down, which is what the autoencoder does.

[00:06:48] Craig Macmillan: So if, if I understand correctly, what we're talking about is , we know that we have the data and we know where wine grapes are grown, different types for different climates. Then we have the climate data in terms of how things may change over time. And then we're creating a prediction of. How those climates change, and then translate that into what we already know about wine grapes.

[00:07:09] Joel Harms: Sort of. Yeah. But in our model for training, we just use the existing ones. So historical climate data and historical grape variety data. Once we have that model trained, we just apply it for new climates that come from like other climate models. So we don't do the climate modeling ourselves, but we extract that information and feed that into it and get the grape varieties output.

[00:07:31] Craig Macmillan: So you look specifically, at least reported on areas that currently are considered too cold for growing a high quality pinot noir or growing wine grapes in general. What did you find out? What Parts of the world might be the new leading Pinot Noir regions.

[00:07:46] Joel Harms: . So that depends a little bit on the exact scenario and how much the climate is supposed to warm. We have like two scenarios is what we looked at. We looked at a 8. 5 scenario and a 2. 6 scenario and going by the 8. 5 scenario, some of the regions that are improving are for example, Western China. And also Southern California, actually, and Quebec, , like Southern California is in Santa Barbara. I guess that's technically Central Coast,

[00:08:17] Craig Macmillan: Yeah, well, that's interesting There's a lot of Pinot Noir in Santa Barbara County in the in the coastal zones Any other regions that popped up?

[00:08:26] Joel Harms: Yeah, a lot of Australia seems to be doing better and like Northern France,

[00:08:31] Craig Macmillan: Yeah pushing it to the north. Did England pop up?

[00:08:35] Joel Harms: England, yes, but England seems to like stay the same in compared to historical. So not like as if it's improving, at least like from this, like rough map that we made. What we want to do is do it a bit more finely. The, this prediction, because we currently just used regions where wine is already grown, but then try to like interpolate just for calculation efficiency. Outward. So like our maps are created not only by the model itself, because that would be too calculation intensive. So for the, for the sake of simplicity, we did it like this, but we're still writing the final paper. So, you know, don't invest just yet, wait a little bit and then,

[00:09:17] Craig Macmillan: I was gonna bring that up. Where should I put my money?

[00:09:19] Joel Harms: Exactly. So don't do that yet. Wait for the final paper and then we will double check everything over. Oh yeah. Arkansas was one that was improving too. Very interestingly. Yeah.

[00:09:28] Craig Macmillan: I was kind of surprised because having talked to guests, many guests from, you know, New York, from Texas, from people who consult in the Southwest Northern California, which can get quite warm. What we've talked about is the question of it getting too hot to grow quality wine grapes.

[00:09:49] You know, wine grapes will grow to tolerate quite high temperatures. So, for instance, the San Joaquin Valley in California, produces a lot of wine grapes. They're not considered to be very high quality compared to coastal zones. So the vines do great and produce good crops and all of that. So there's concern that areas that have been kind of in the sweet spot, kind of in the, we call it the Goldilocks phenomenon where climate, soil, time, everything just all kind of fits together.

[00:10:12] It sounds like this idea would be applicable to predicting what areas might become too warm for high quality wine

[00:10:19] Joel Harms: Yes. Yes. It's definitely the case. Yes. And in our maps. You can see both at the same time because it sees like relative change, positive, relative change to, to negative. Some areas that look like they're not going to do so well in the future or less good in the future, even though they're like really good right now is like Oregon, unfortunately.

[00:10:39] And the Azores or Northern Spain, even in Eastern Europe, a lot of areas. Seem to be warming up like in Romania at the coast. Not necessarily just the warming up part, but also because we consider 16 different climate variables, it could be the warming up part, but it could also be, you know, like the precipitation changing things like that, you know.

[00:10:59] Craig Macmillan: You said 16 variables, we talked, you got temperature, you got precipitation, what, what are some of the others?

[00:11:04] Joel Harms: Yeah, we got the growing degree days, the winter index, we got the Huggins index, we have radiation. Diurnal temperature range, the annual average temperature, for the precipitation, we have it like a three different scales, in the harvest month over the growing season and also throughout the whole year same for the temperature. And then we have the, growing indexes

[00:11:26] Craig Macmillan: do you have plans to do this kind of thing again? Or publish additional papers from the work you've already done, because I think, it sounds like you've got a lot of interesting findings,

[00:11:35] Joel Harms: Oh yeah. Yeah. The results only came in like right before the conference. We're still analyzing everything, writing everything. So the first thing that's coming up is a paper just on , how did we build the model and like all the validations and does it make sense with like expert classifications of how experts classify suitability for grapevines and things like that in the past to see if. That lines up as it should yeah, and then after that we'll publish some of these predictions and what we can learn from these and more detailed than how we did it right now where, most of it's like interpolated because we couldn't predict for every location, so like we predicted for some locations and interpolated. Just for computational efficiency, I guess, but you know, we're, we're getting there. Unfortunately, academia is quite you know, a slow profession. takes a lot of time.

[00:12:24] Craig Macmillan: Yes, yes it does. And then getting it published takes a lot of time with reviews and whatnot. And so I just want to put a time stamp on this. This is being recorded in October of 2024. So, Give it some months, at least several, several, several, several. But it's exciting. This stuff's coming out. It'll be in, be in the literature. That's really, really great.

[00:12:43] Joel Harms: And soon what we're trying to do is also release like a tool or something that, you know, where people can input their location and we can, our climate data, like call out the climate data and see what, what some of the predictions would be. Yeah.

[00:12:57] Craig Macmillan: Oh, that's neat.

[00:12:59] Joel Harms: I might've done that for Niner Vineyards just now to see, to see what, what's a suitable there, but only the current ones.

[00:13:08] So I mean, it's kind of is exactly what you're growing.

[00:13:10] Craig Macmillan: Funny. You should mention that. There is a a website called CalAdapt that allows you to put in some ranges and some variables specific to your location, you put your location in, and then there's a number of different models that you can run. Some are very conservative, some are not in terms of what the predictions are for climate change globally.

[00:13:31] And then gives you a nice report on what the average temperature change might be in degrees Fahrenheit or Celsius also takes a stab at precipitation, although I talked to somebody who was connected to that and they said the precipitation is always kind of questionable. And also looks at things like heat waves, how many heat waves days over 100 or days over 95, you might expect because those can be quite fluctuating.

[00:13:55] damaging. Even, even though vines can tolerate heat, if they're not acclimated, getting these big stretches of over a hundred, for instance, can be kind of stressful. I did that and kind of looked at it myself and thought, huh, I wonder if we had better, more, um, detailed information, what that might look like.

[00:14:12] Another tool that was mentioned that you used was a deep coupled auto incoder networks. What are those?

[00:14:18] Joel Harms: So that was what I described earlier, like these component models , where we have a. The encoder and decoder part, the input part is the , encoder and the output part is the decoder. And in the middle of these we have a latent space and then the coupled part means that we're having multiple of these that share their latent space.

[00:14:38] So that's , where we're tying them together so that we can input either climate or grapes and get as outputs either climates or grapes. So it's like very, very flexible in that way and so I quite like that. And it turns out it does better than even some more traditional approaches where you just feed in climate and get out grapes like from a neural network or something like that.

[00:14:59] Just like a neural network, because we have technically like four neural networks and all of them have three layers. So that's three layers or more. And so that's what makes them deep.

[00:15:08] Craig Macmillan: Got it.

[00:15:09] Is this your primary work as a PhD student?

[00:15:13] Joel Harms: Well, as a PhD student, I'm still working on modeling. But not so much with grapevines, unfortunately. I'm looking at still climate models. How can we adapt for example, now we're looking more at the Caribbean. There's flooding issues. Particularly in Guyana. And so we're trying to, you know, help maybe the government to plan land use better in order to avoid, you know, critical areas being flooded, agricultural land being flooded and these type of things.

[00:15:41] So it's more looking at flooding modeling, there's definitely some overlap in that sort of work, it's definitely still like in the area of using data science to help decision making which is the overall theme of this work.

[00:15:55] Craig Macmillan: Yeah, and that was something that also came up in my little mini project was the potential for massive storms and also the potential for drought. Which, wasn't part of your work at this stage. Is that something that you would be able to find a way of including in your modeling that might give you some idea of how things might change?

[00:16:15] And it's specifically what I'm thinking of is Cyclone Gabriel, I believe it was called, Gabriella just devastated parts of New Zealand. And raised a lot of concern about how, you know, when we were in these coastal zones, we go, Oh, yes, it's mild. It's great. But we're right near the ocean.

[00:16:33] Right. And in October between 24, we've seen a very active hurricane season in the Caribbean and on the East coast and the Gulf. Do you think there's potential for this kind of thing to give us more of a heads up about what might be coming our way in terms of massive storm events? Cause that might affect how and what I do.

[00:16:52] Joel Harms: I guess this wouldn't depend really on the grape variety itself. That would be more like a citing issue, right? Like where do you plant?

[00:16:58] That's what we're looking at now with the like flooding mapping if there is a storm, where does the water collect? Which roads are cut off? Or, I mean, I guess in the case of vineyards, you could look at like, what would be the likely damage would there be now saltwater maybe even if you're depending on where you are. That's definitely something to look at.

[00:17:17] All you need is sufficient, like past data points. So you can calibrate your models and then. You know, look at different future scenarios and what will be important to for the future is to look at what's kind of the certainty of these predictions, right? Like, what are your error margins? What's your confidence interval?

[00:17:33] Because that might drastically alter your decisions. If it says, oh, it's probably not going to be too bad, but you're very uncertain about that, then you're probably going to take some more precautions than, you know, not because usually now we have A lot of models where their prediction is very, like is deterministic.

[00:17:50] So they say, this is how it will be. And it's hard to tell where, you know, where those margins are of error, which is something to look at in the future for sure.

[00:18:01] Craig Macmillan: Yeah, that is a challenge in the the model that I did for a Paso Robles vineyard Precipitation didn't really change very much which I was surprised by so it wasn't gonna become like a drought area completely but the potential ranged from five inches of rain a year to 60 inches of rain a year, which is why I was asking about these massive storms.

[00:18:21] Maybe our averages, continuous to what we have now, but it may be a bunch of craziness year to year around that. And I think that is interesting and useful to know. So you prepare for it.

[00:18:34] Joel Harms: that's something people are looking at, I think cause you can use some models to calculate sort of new climate indices. To see like from daily data train, like new climate indices to see these big storm events and things like that, and maybe incorporate that. That could help, , maybe with that sort of analysis of where even if it's the same average, the index is different because it measures something else.

[00:18:59] Yes, I wouldn't know what they're called, but yes, I believe this already exists and is being improved. .

[00:19:05] Craig Macmillan: Yeah. Yeah. With your experience so far, what do you see? Because everybody's talking about this. It's like the future in a world of artificial intelligence and this and that. In this particular area where you're, you're tying one set of variables to climate variables and also to historical weather.

[00:19:23] In the big picture, beyond just wine grapes, but in the big picture, any topic, where do you see this kind of work going? You touched on it a little bit, when you close your eyes and open your mind what does the future look like? What, kind of tools are we going to have and what kind of things are we going to be able to find out?

[00:19:38] Joel Harms: Yeah, that's interesting. I think it, it really depends on the data we have available and it looks like we'll have more and more data available.

[00:19:47] So better disease models, location specific disease models to plan spray schedules better and things like that, they seem to be coming. I think I've seen parts of that already from some companies rolling out.

[00:20:00] It's all about kind of the creatively using the data that you have available, because a lot of like my data, for example, that I used for this. This isn't necessarily new data, right? This comes from the University of Adelaide who collects where, which grape varieties are grown all over the world.

[00:20:17] And then just historical, climate data. It's not very new, but just to put these together in a meaningful way with AI, that's going to be the challenge. And then also to test, is this reliable or not? Because you could theoretically predict almost anything, but then you need to check, is it just correlation?

[00:20:39] Am I taking all the important variables into account? And we're developing AI very, very fast. But maybe we need to spend a bit more time, you know, trying to validate it, trying to see how robust it is, which is a major challenge, especially with these complicated models, because, I heard about this example.

[00:20:57] Where in the past, for some self driving cars, their AI that recognized stop signs could be tricked if there was a sticker on the stop sign, and it would ignore the stop sign. Even though there's not a big difference, but you can't test for, you know, all of these cases, what might happen. And that's kind of the same for, , what we are doing.

[00:21:17] So improving the testing, that would be, I think, a major A major goal to make sure it's robust and reliable or that it tells you how, how certain it is, you know, then at least you can deal with it, you know, and not just make a decision off of that. Yeah,

[00:21:29] Craig Macmillan: Yeah. What the level of uncertainty is. That's always the getcha.

[00:21:33] Joel Harms: yes,

[00:21:34] Craig Macmillan: That's always the hard part. If you had one thing that you would tell growers on this topic, what would it be? Mm

[00:21:43] Joel Harms: Specifically for my models, it would be to take the current results with a grain of salt. And then to sort of use this to, narrow down like a selection of grapes and to still run tests and things like that. Cause it's regional data, right? It's not going to tell you exactly what you should grow in your location.

[00:22:02] Cause it's, you know, the weather data is based on four to 50 kilometers around you. You know, that's where we're like assembling the data from.

[00:22:10] Craig Macmillan: that a 50 kilometer quadrant?

[00:22:12] Joel Harms: yes. Yeah.

[00:22:13] Craig Macmillan: Yep. Okay. Gotcha.

[00:22:14] Joel Harms: Yes, exactly. So this tool is mainly used or useful if you use it to like pre select some varieties so you can see what might be good, you know, and then decide for yourself what you want.

[00:22:27] The take home message is like, it's not supposed to take away grape growing experts and things like that, or replace them in any way, but it's supposed to like support it because. There's so many grape varieties and if climate regions or like regions where we're growing grapes are changing, where the climate is changing, we want to get the best choice.

[00:22:47] And so we should probably look at all of them, all of our available options and see what we can do. It will narrow it down for you. And then, you know, you'll still have to see what works exactly for you. What wine do you want to produce? I mean, it doesn't take that into account, right? It just gives you what probably would grow well here.

[00:23:03] Craig Macmillan: .

[00:23:03] Yeah, then I think that there's going to be a future also in bringing in some either hybrid varieties or varieties that are not terribly well known. I've talked to people from Texas and from Michigan Pennsylvania, where the traditional vinifera only varieties don't do pretty well. Terribly well, often because of cold hardiness because of cold winters, they don't handle it, but there's hybrids that do great and make interesting wine.

[00:23:27] And I think that would be an interesting thing to include in a model or if it came out kind of like the winner was something we don't normally

[00:23:33] Joel Harms: Right. Usually we have a lot of hybrids in this because we have 1, 300 varieties.

[00:23:39] Craig Macmillan: wow. Oh, I didn't realize that.

[00:23:41] Joel Harms: so I think we have most of the. commercially used grape varieties, like in all aspects.

[00:23:48] Craig Macmillan: yeah, probably, probably.

[00:23:49] Joel Harms: Yeah. So it's quite, quite far ranging. We only excluded some where it was never more than 1 percent of any region, because then like our model couldn't really learn what this grape variety needs.

[00:24:00] Right. Because it's like too small, even in the largest region where it we cut those out. So, cause else we would have 1700. But then like the 1300 that actually get used commercially at a significant scale. Those we have. The model is actually built like we have a suitability index.

[00:24:18] But we're still trying to, , fine adjust so that we can rank not just what's popular and like how much will grow. Cause then you'll always get, you know, the top, the top 10 will look very similar for any region. But then through the suitability index, we actually get a lot of these smaller varieties that would fit very well also ranked in the top 10 or in the top 50 of varieties.

[00:24:41] Craig Macmillan: They've mentioned fine tuning the model at this point. Is this particular project or this particular model, is this gonna continue on into the future? It sounds you have ideas for improvements. Is this number one gonna continue on into the future and is there gonna come a point when This will be available for the industry, industries internationally to do their own trials.

[00:25:03] Joel Harms: Yes, I think so. So I think when we're publishing the paper latest at that point, we'll have the tool set up where people can try it out, put in, in their location. And I guess we're publishing the methodology. So you could build like a version of this yourself. It's not too crazy. Probably code will be published too.

[00:25:24] So, you know, you could build this yourself if you wanted to, or you could just use the models we have trained already. Okay. And just apply them to your case. That's what the tool is for. . Right now it's like all code based. So like, it's not, not so easy where you just, drop your pin, like where you're at and then it gives you some predictions, , that's what we're aiming for.

[00:25:44] Craig Macmillan: Fantastic. So our guest today has been Joel Harms. He is a PhD student in the Department of Bioresource Engineering at McGill. University. Thanks so much for being on the podcast. This is really fascinating. I'm really looking forward to how this work progresses. And I think it's very eyeopening for us.

[00:26:01] Again, you know, one of the things I thought was fascinating is I've had all these conversations about areas that would no longer be suitable, but a flip on it and say, well, areas that might be suitable in the future. I hadn't thought of that.

[00:26:12] Joel Harms: Why not? You

[00:26:13] Craig Macmillan: why not? You know, that's, that's, that's a very interesting question, and it applies to other crops as well.

[00:26:18] I just had never really thought about it like that. You know, maybe you can grow oranges in Iowa at some point.

[00:26:23] Joel Harms: That, that would be nice. I guess.

[00:26:25] Craig Macmillan: maybe

[00:26:26] Joel Harms: maybe see.

[00:26:28] Craig Macmillan: we'll see. We'll see. You never know. Anyway, Joel, thanks for being on the podcast. I appreciate it.

[00:26:33] Beth Vukmanic: Thank you for listening. Today's podcast was brought to you by Cal West Rain. Since 1989, Cal West Rain has served growers on California's Central Coast and the San Joaquin Valley. As a locally owned, full line irrigation and pump company, they offer design and construction experience in all types of low volume irrigation systems, whether they're for vines, trees, or row crops.

[00:27:03] In addition, CalWestRain offers a full range of pumps and pump services, plus expertise in automation systems, filtration systems, electrical service, maintenance and repairs, equipment rental, and a fully stocked parts department. Learn more at CalWestRain. com.

[00:27:23] Make sure you check out the show notes for links to Joel, his research articles, plus sustainable wine growing podcast episode 207. Managing Catastrophic Loss in Vineyards, Lessons from Cyclone Gabriel in New Zealand. If you liked this show, do us a big favor by sharing it with a friend, subscribing, and leaving us a review.

[00:27:44] You can find all of the podcasts at vineyardteam.org/podcast, and you can reach us at podcast at vineyardteam.org. Until next time, this is Sustainable Wine Growing with the Vineyard Team.

 

Nearly perfect transcription by Descript