Faculty Interviews

Monika Calef, Associate Professor of Physical Geography

In August 2023, the PBRC Director had the opportunity to talk with Monika Calef, Associate Professor of Physical Geography and the current director of Environmental Studies, about her recent research funded in part by the PBRC, in particular, her recent publication in Forests: “Predicting the Unpredictable: Predicting Landcover in Boreal Alaska and the Yukon including Succession and Wildfire Potential.” (This interview has been edited and condensed for clarity.)

J P Kehlen
Could you tell me about the origins of this project and how it came about. For many faculty, there’s usually moment spent reading or thinking and all of a sudden we think, what would happen if I asked this question? So was there such a moment for your work?

Monika Calef
This paper started with an email from somebody asking me if I can give her my vegetation layer map that I created many years ago when I was a postdoc. I used logistic regression to create a vegetation map of the Western Arctic — that’s Alaska and Western Canada. But it was at a one kilometer resolution, one vegetation type for each kilometer. So it was very simplified, and she wanted a much more detailed one. The one I had created wasn’t very accurate, and the resolution wouldn’t work for her; she wanted it at 30 meters. And you can’t just go from one kilometer to 30 meters. I mean, you’re going to end up with garbage.

JPK
That’s three orders of magnitude of difference, right?

MC
Exactly. So then we decided to basically work on it together. She just asked me, hey, can you send me that layer and then four years later, I actually had a final product. It’s been very, very slow going. It was immensely complicated.

JPK
This is one of those amazing aspects about academic life that sometimes isn’t always apparent, that a chance email or encounter between two people in different fields can throw off a mental spark, and from that, an entire new kind of analysis can emerge.

MC
That’s true. It’s funny because her big project that looking at land cover change in three locations in the western Arctic. And she worked with economists and sociologists to go to villages there in order to tell them about future scenarios and how they can address the future areas that are going to be prone to wildfire or other damage. It’s really all about working with communities to help them address better preparedness for the future, to assess risk and vulnerability, and then to come up with better scenarios. The vegetation map was just one little basis that then fed into all the other questions. And so this one small thing that was just going to be sort of a spark turned into this huge project that four of us worked on for four years.

JPK
So something that seemed not trivial but relatively unimportant in one study then became the central aspect of a new study altogether.

MC
Yes, kind of a spin off. And I was happy to work with them because it was a great opportunity for me to work with a big research group. Then I quickly realized that I’m wading in over my head. And so I asked Anna Varvak to help because I’ve worked with her on to other papers, doing statistical analysis because it all comes down to statistics, and it’s all spatial. So it violates basically all the normal assumptions of statistics, and you need very advanced statistical models, and then trying to find all the data sets. It was really a lot of my collaborator Jen’s ideas and Anna’s analysis and I’m kind of the go-between.

JPK
And here the go-between is terribly important for translating data and information from one field to another in a way that is useful for them both.

MC
Yes, and because I had worked with Anna previously, I knew how to talk to her with a math background and from us with the science background, I could give her data and she could analyze it, but some of her analyses didn’t make any sense from an environmental perspective because she didn’t understand what she’s looking at.

JPK
You make it sound like serving as an interpreter between different languages. That sounds like the definition of interdisciplinary.

MC
It’s completely interdisciplinary. And that’s why it’s difficult to explain exactly what we did, because there are so many different arguments that create the entire project.

JPK
Could you give me a simple explanation of land cover and why it is so important to this study? It’s clearly a term that’s fundamental to your research.

MC
Land cover is basically when you look at the landscape, what does it look like? Is it trees, is it tundra, is it water? What is the surface like — what’s the vegetation, essentially — or does it not have vegetation? And that has important implications in terms of climate, fire, permafrost. And those are the things we were looking at. So the permafrost is usually under black spruce trees, because they are the only trees that can grow on permafrost. That means the ground has been frozen since the last ice age, right? So it’s been frozen for thousands of years and still that’s the only tree they can grow on it; when you wander around the landscape and you see black spruce, then you know it’s going to be frozen underneath it. It gives you some idea of what the soil is, what the moisture is, is there acidic water or not under it, what kind of rock is underneath it. So just by looking at the vegetation you have a very good idea of sort of the larger environmental picture of that place.

JPK
It’s amazing to hear you say this, because I just heard you define land cover in narrative terms — that every different type of land cover tells us a story about not just its location or what grows there, but about the entire history of that specific spot.

MC
Yes, that’s essentially what happens and that’s why we want to know, how does the land cover change over time. And that was really the going question for this project: from 1980 to 2060, what does the vegetation and land cover look like in Fairbanks, Anchorage, and in Whitehorse? Because that gives us an idea of how will the vegetation and the permafrost change, and how does the risk of fire change, because fire only burns mostly in black spruce. So if you have a house among black spruce, it’s going to burn. [The analysis] tells us about vulnerability to fire from these soil changes. And that’s what we were interested in because that’s the information communities need, so they can figure out how to plan for the future. If you’re going put people there, then they need to be able to escape and you can’t just have one road, you need to have multiple routes of egress — this kind of scenario planning is critical for the communities; they need to have kind of a future landscape where they can see what the future might look like. That was the driving factor behind this project.

JPK
Paraphrasing what you said, looking at land cover through time can reconstruct data beyond the present as a way to predict what is going to happen, which is incredibly useful from many different perspectives for the communities and government in those locations.

MC
But it’s so hard to do, which is why the title is “Predicting the Unpredictable,” because we can’t know the future. Climate change is messing up everything; we know it’s getting warmer, we know it’s getting drier in those areas. That means things are going to burn more but none of this change is linear; none of it is perfectly predictable. Your d is some ridiculous number because you have wet years, you have dry years, you have super-warm years, then you have cooler years like this one: it rained all summer and nothing burned. So a major problem is, how do you predict the future if there’s so much noise in the in the data.

JPK
Which connects to the varying degrees of reliability discussed in your paper.

MC
Yes, exactly. There are large climate cycles; in theory, every 30 years it gets drier and then it gets cooler again and that influences vegetation and fire. But there are other cycles that we don’t even of know yet. So this is where climate science is just starting to discover these connections, because in order to figure out 30-year cycles, you need 100 years of data. Our satellites have only been around since the 1980s, so we don’t have that long of a record. That’s why these patterns are just now slowly starting to emerge with supercomputers and high resolution satellite data, and we’re starting to see these things because they were not immediately obvious. That’s why we started in 1980 to see how has the vegetation changed in that time, what does the climate look like, what does the fire situation look like. And what we found is that you already see extreme fire years, such as 2004, when my son was born, everything burned, and we were trying to leave just to get away from the bad air. You have years like that every couple of decades. But then in-between, it’s not so bad — like this year, it rained all summer. That’s why we decided to do decadal analysis — we throw 10 years together and take some kind of average because there is much variability. Then that gives us a kind of estimate: “Okay, we have the sort of decades where you don’t have a lot of fire and then you have a decade where you have these huge fire years, maybe just one, but it raises sort of that area burned tremendously by a factor of one or two.” That’s why we decided we’re going to use it for predicting the future. And then there’s the whole feedback from the vegetation. Once you burn all the black spruce in the landscape, you don’t have any fuel left, so it’s not going to burn and that’s what all the other models are showing — that the vegetation is changing. The black spruce is vanishing from the landscape. Of course it can grow back but it needs 60-70 years. But if the fire comes back too soon, then you don’t have spruce trees coming back. So you’re changing the landscape, and that’s the hope for the future, that once you burned all the black spruce, the fires will stop. But now it’s a bad time for these communities because the fires will just keep going up. And of course, that keeps feeding back into the carbon cycle; you will have more carbon released from these forests and the permafrost, which is a huge carbon sink and methane from the bogs, which is even more powerful. So then you have feedback that is just going to make it hotter and drier and you’ll get more fires and then you have even more feedback.

JPK
Please say more about how you use statistics in this project.

MC
Well, we have different study sites, and some of the data layers cover the whole study site, and some of them don’t, because it’s so difficult to find data for Alaska. That’s why we’re trying this study, because there was a very intense Western Arctic exploration from 1984 to 2014, which was our basis. We also found other types of data, each with its own issues of interpretation. For instance, I had data for three different vegetation layers in and around Anchorage. They were created at different resolutions and I was trying to use them to verify each other, but they are completely different — they have different definitions for land cover types; one of them says this is a spruce forest, the other one says this is a shrub layer, and so they’re all really, really bad. And that’s a problem because you have satellites flying over the landscape taking pictures, but they’re not going to be that accurate, right? I mean, the algorithms are getting better, but the precision’s still not there. So a lot of the statistics I use correlates datasets to see: do they correspond with each other? For example, data for tree height or different soil parameters, such as how much moisture is in the soil or the pH of the soil water. Sometimes you find that some of these data are showing you kind of the same thing. So you can drop some parameters out; you don’t need to have 20 parameters if 10 of them are kind of the same thing. Then you look at correlation: if just one goes up do the other ones go up too, and then can I just not use them? So this is back to my knowledge of the landscape. Understanding what these things mean and the statistics of okay, if they have an R-squared of close to one. Obviously, it’s not going to add any new information if I add this one.

JPK
So you’re taking two different modes of analysis — environmental science and statistics— and find the place where they meet —

MC
Right. And, understanding that the metrics are nearly the same thing environmentally speaking, the statistics show me that I don’t need all the parameters.

JPK
Then you’re using these statistics as a way to translate numbers into a functional understanding of environment.

MC
Yes, exactly. That’s a good way of saying it.

JPK
I remember that Anna once told me that people are afraid of statistics because they think, it’s just math, math, and more math, but when we think about the purpose of statistics, which is to create a better understanding of the real world, most people find statistics then becomes incredibly interesting.

MC
My life is all about patterns. As a geographer, I’m trained to detect spatial patterns so I can make a map and look at it. I don’t necessarily see the patterns on that map. But if I run statistics on the data, I can then see the patterns because its statistics will show me if there’s a pattern or not. And then I have to go back to the map and think, yeah, I can kind of sort of see that, or overlay different parameters and then it’s starting to crystallize for me.

JPK
It’s interesting to hear you talk about patterns, because, as a philologist, I’m also always looking for patterns in language through history and culture as a means of getting a clearer insight into “what did this person mean by using these words?” And in almost every field of scholarship — natural science, social sciences, humanities — we’re all looking for patterns and trying to figure out how to interpret them, even if the ways we describe those patterns or analyze them are often startlingly different.

MC
That’s true, and for me, it’s even more important to ask, which statistic could I use to tease out that pattern, or how could I overlay these data or change them? So Anna did a lot of this work. She basically converted data logarithmically, like the natural log, because otherwise there’s too much noise in the data; the fire could be one acre, or it could be a million acres. If you put that in linear regression, it’s just going to be blowing up. By transforming the data, you normalize it by dividing it by the highest number, so that you rescale everything to 100. Then it becomes a little better to start overlaying the data and analyzing them. So it’s little data tricks like that that I’ve now learned from doing this study.

JPK
That sounds like you became a data detective.

MC
Yeah. But that’s where I get excited. And I love data! So that was the prep work figuring out which datasets to use, where to find them, which ones work for us. And then it was a question of, okay, I want to predict the vegetation. I have these data sets. How do I put that together so it gives me a future vegetation layer? And that’s where Anna came up with a computer machine learning model, which was able to take the data we had and to interpolate it. So you train the computer and you do different testing to make sure you will have the best statistical models, different ones for each of our time periods and locations. And then, we run our statistics into our fancy model to see what it does — something would have been impossible to do without automation. It is an ensemble learning algorithm that combines prediction for multiple decision trees; it builds a decision tree based on the data, and then it builds another one and refines the previous one — every time it runs through, it improves on itself. You run this over and over and over —

JPK
Which is totally different from just using randomization —

MC
Exactly. That’s something we couldn’t even do 10 years ago, because we didn’t have the capacity for all these statistics. Now that the models keep getting fancier and the computers keep getting faster, we can do this.

JPK
Machine learning isn’t something to be afraid of, in other words,

MC
No, not at all. We couldn’t do this study without it, because it’s so complex.

JPK
One last question. At the end of your paper, you write, “Our relatively simple modeling approach could serve as a blueprint for similar efforts elsewhere.” What are the possible implications of that sentence? Because it suggests that not only your findings but the methods themselves may provide future paths for other researchers in a variety of fields.

MC
We did use a very simple model. Other models that predict the future are super-complex ones that people, some of whom I’ve worked with, have been working on for years and years. My postdoc was with someone who has a model that runs on multiple platforms and he’s had full-time programmers creating these models that calculate all the detail of the landscape; you’ve got climate, you’ve got the vegetation and all the soil interactions, which is critical because you need to know how the water cycles. You need to know the soil types. You need to know topography. That’s dictates how vegetation and fire interact with each other, so those models are hyper-complex. But our model is really a simple statistical model, and it’s free. You can literally download it with our data; everything we used is available online. It’s all free government data. So compared to the existing models now in use that are so extremely complex and, of course, owned by somebody who has spent 10 years building them, ours is relatively simple and open-source.

JPK
A way to look at complex data without oversimplifying and still make that useful for a variety of different fields or types of research.

MC
Yes, and there are multiple uses because everything’s changing. I mean, climate is changing everywhere. I’m thinking that someone who doesn’t have a fancy, expensive computer models at their disposal could download our script, and they can modify it, get some kind of data that they can find for their area, and then use this approach in general.

JPK
An outcome that’s beneficial for your own field and for other researchers.

MC
We spent four years figuring it out, but now that we figured out how to make it work, you can modify it to whatever will work for your study, so it’s incredibly flexible, and you don’t need that proprietary model that’s going to then run in some kind of computer ensemble for a month or more. 

JPK
Congratulations on your project, and thanks for giving us more background about your research — and for sharing your enthusiasm about data.

MC
I love talking about statistics! And thanks to the PBRC for the grant that helped that fund our research.