Last updated: October 21, 2023
Article
Podcast 139: Southern Pine Beetles and Archeological Site Modeling.
A New Project On… Beetles?
NPS, NCPTT
Tad Britt: Tad Britt, and our guest today is:
Zac Selden: Zac Selden, from the Heritage Research Center at Stephen F. Austin State University.
Sadie Schoeffler Whitehurst: Thank you, guys, alright. Today we're going to talk a little bit about Zac's grant through NCPTT. He's going to tell us a little bit about his research: Archaeological Site Modeling for the Sabine National Forest in Advance Preparation of the Southern Pine Beetle. If you would, tell us a little bit about your project.
Zac Selden: I think this really all spun up after I met Tad and went to a predictive modeling seminar up at the University of Arkansas. There, after [the seminar, I] met with the Forest Service folks and we had discussed different ways of modeling the Davy Crockett National Forest at that time. While all of that was spinning up, I also had been in conversation with Tad about what's the best way to go about predictive modeling, what are the best techniques, what can we use, and what is the cutting edge. We worked then to use a bibliometric survey to identify the tools and techniques that would kind of give us the edge we needed to be competitive. Identified the software and the methods that we wanted to use, and once we had all of that ready, [we] moved the collections from the national forests and grasslands in Texas to my lab where they were documented, and we've been chasing predictive modeling kind of ever since. This particular project was designed knowing that we have a native species that will impact the forest and the forest responds to that impact in a particular way. However, to date, they've been managing it mostly through informed guesswork, and if we had a model like the one we're trying to build, that would help them to better manage those properties and hopefully make a more data-driven decision in terms of how they manage those situations where potential cultural heritage resources are impacted.
Southern Pine Beetles and Archaeology
Image courtesy of Zac Selden
Zac Selden: It eats trees [laughter]. It heavily impacts pine, as you can imagine, and because of climate change, the range of that beetle is expanding. Where it was this traditional American southeastern pest that you know occurs, it kind of flash mobs a forest and then decides, “alright, well, we're going to move west or we're going to move east,” wherever it starts that particular season. Well, now they're moving west. They hit Mississippi, I think 2-3 years ago and are in Louisiana now. The goal is to have the Sabine national forests model up and rolling before they are impacted by the southern pine beetle. They have now spread upward to northern New York and now New Jersey. It has become a much larger issue and hopefully, at least it's our hope, that once we have the model built in the code and everything is ready, we can make that exportable to other forests, other properties where they could use it.
Predictive Modeling
Sadie Schoeffler Whitehurst: How does the process of building a predictive model work?
Zac Selden: So, it begins with sites. The site information is just XY data. We wanted to know what variables we should use in that model, and that took several conversations, not only back and forth with Tad originally, but then with the National Forests and Grasslands personnel. The personnel that were on the Davy Crockett helped us to pick what they saw as environmental variables that correlated with site locations. For that particular piece, we were looking at proximity to water, elevation, soil, and vegetation. For the Davy Crockett model, we threw absolutely everything at it just to see what's stuck, and there are ways to get the model through using a bootstrap method. It gives us a means of identifying those variables that most closely articulate with your sites. When we use those that were most impactful as a means of really kind of trying to squeeze as much accurate information out of the model as possible without going too far if you go too far, it's called overfitting, so you get these high accuracy levels, but it's not generalizable. The goal with the Davy Crockett was design, research, and development and kind of really tweak and learn, as a group. We wanted it to be useful for predicting prehistoric sites generally, and historic sites, and then they were going to use a model for all sites for more CRM-based endeavors. For me though, so I guess it's twofold. You know, any kind of predictive model is going to have applications in applied research, which is what we're doing there. My own kind of selfish motivation was to look at how these resources change through time for the Caddo period. I haven't made a whole lot of headway there admittedly, but it is something that's on my radar, particularly with this model, because I think that it's really the first test of how generalizable DE David Crockett model really is.
Tad Britt: Zac, can you elaborate on the maximum entropy model approach that you take?
Zac Selden: So, we started with a GUI, a graphical user interface, where everything was point and click, and we have since evolved since Maxent went open source and is now available through R. We wanted the code to be completely exportable and user-friendly to the extent possible so that folks can plug and play that model anywhere that they have these kinds of resources and these kinds of risks. I think the crux of that would then be what variables you plug in, it's going to differ in every location, but the algorithm that we're using, it's a machine learning procedure. It gives us a means of using 80 or 70% of the sample to really define our predictions, and then the remaining 20 or 30% to then test those predictions and get a good idea about how accurate they are. We're using these to make predictions for the forest. One of the side projects that I've been kind of slowly whittling away on is, is it possible to then look at maybe something as simple as mounds? Where do we have mounds that are occurring in the woodland period, deformative early cattle period, middle, then late historic, and how does land use change? Do we see the intensification of populations around certain resources, and do we see the movement of people across space? We absolutely do. I think that raises a whole bunch of questions about if they're moving, what are they taking with them? What new tools are they developing along the way and how do those tools evolve as well? Pairing this I think with other advanced tools, maybe geometric morphometrics, can be useful in looking at not only the evolution of how folks use space, but maybe the evolution of the tools that they took with them, and the ideas that they took with them. I think there's a there's a lot of different things that you could plug into it, and it could be anybody's passion project. If you wanted to look at Munsell colors, if you wanted to look at the geology, maybe the correlation of a certain environmental variable or a certain lithic. Do you have obsidian occurring in one place and not in another, and then did those folks move across that boundary? We definitely have boundaries in the Caddo region that we know exist, but mapping movement has been a challenge. Having a tool like this where we could not only look at and kind of better understand that movement but also, in the end, get a model that would tell us those regions where the recipe is right for other resources to exist, gives us a means of better managing maybe those invisible resources that perhaps we haven't discovered yet. The ultimate goal: I guess the production of reliable knowledge would have to be number one, and for that knowledge to be useful for the folks that come behind to be able to use it and to maybe build upon it to better understand guesswork that we're putting into understanding this patchwork that is the past. I would like to say thanks to NCPTT for helping by funding the project. I do need to give a shout-out to the National Forest and Grasslands in Texas who funded the fieldwork component of all of this. We're not only building the model, but we're going out to test those models. That is exciting and fun and hopefully will help to mitigate some of the survey bias through the years.
Sadie Schoeffler Whitehurst: Absolutely! Ok well, thank you, Zac.
Zac Selden: Thank you.
Tad Britt: Thanks.
Zac Selden: So, it begins with sites. The site information is just XY data. We wanted to know what variables we should use in that model, and that took several conversations, not only back and forth with Tad originally, but then with the National Forests and Grasslands personnel. The personnel that were on the Davy Crockett helped us to pick what they saw as environmental variables that correlated with site locations. For that particular piece, we were looking at proximity to water, elevation, soil, and vegetation. For the Davy Crockett model, we threw absolutely everything at it just to see what's stuck, and there are ways to get the model through using a bootstrap method. It gives us a means of identifying those variables that most closely articulate with your sites. When we use those that were most impactful as a means of really kind of trying to squeeze as much accurate information out of the model as possible without going too far if you go too far, it's called overfitting, so you get these high accuracy levels, but it's not generalizable. The goal with the Davy Crockett was design, research, and development and kind of really tweak and learn, as a group. We wanted it to be useful for predicting prehistoric sites generally, and historic sites, and then they were going to use a model for all sites for more CRM-based endeavors. For me though, so I guess it's twofold. You know, any kind of predictive model is going to have applications in applied research, which is what we're doing there. My own kind of selfish motivation was to look at how these resources change through time for the Caddo period. I haven't made a whole lot of headway there admittedly, but it is something that's on my radar, particularly with this model, because I think that it's really the first test of how generalizable DE David Crockett model really is.
Tad Britt: Zac, can you elaborate on the maximum entropy model approach that you take?
Zac Selden: So, we started with a GUI, a graphical user interface, where everything was point and click, and we have since evolved since Maxent went open source and is now available through R. We wanted the code to be completely exportable and user-friendly to the extent possible so that folks can plug and play that model anywhere that they have these kinds of resources and these kinds of risks. I think the crux of that would then be what variables you plug in, it's going to differ in every location, but the algorithm that we're using, it's a machine learning procedure. It gives us a means of using 80 or 70% of the sample to really define our predictions, and then the remaining 20 or 30% to then test those predictions and get a good idea about how accurate they are. We're using these to make predictions for the forest. One of the side projects that I've been kind of slowly whittling away on is, is it possible to then look at maybe something as simple as mounds? Where do we have mounds that are occurring in the woodland period, deformative early cattle period, middle, then late historic, and how does land use change? Do we see the intensification of populations around certain resources, and do we see the movement of people across space? We absolutely do. I think that raises a whole bunch of questions about if they're moving, what are they taking with them? What new tools are they developing along the way and how do those tools evolve as well? Pairing this I think with other advanced tools, maybe geometric morphometrics, can be useful in looking at not only the evolution of how folks use space, but maybe the evolution of the tools that they took with them, and the ideas that they took with them. I think there's a there's a lot of different things that you could plug into it, and it could be anybody's passion project. If you wanted to look at Munsell colors, if you wanted to look at the geology, maybe the correlation of a certain environmental variable or a certain lithic. Do you have obsidian occurring in one place and not in another, and then did those folks move across that boundary? We definitely have boundaries in the Caddo region that we know exist, but mapping movement has been a challenge. Having a tool like this where we could not only look at and kind of better understand that movement but also, in the end, get a model that would tell us those regions where the recipe is right for other resources to exist, gives us a means of better managing maybe those invisible resources that perhaps we haven't discovered yet. The ultimate goal: I guess the production of reliable knowledge would have to be number one, and for that knowledge to be useful for the folks that come behind to be able to use it and to maybe build upon it to better understand guesswork that we're putting into understanding this patchwork that is the past. I would like to say thanks to NCPTT for helping by funding the project. I do need to give a shout-out to the National Forest and Grasslands in Texas who funded the fieldwork component of all of this. We're not only building the model, but we're going out to test those models. That is exciting and fun and hopefully will help to mitigate some of the survey bias through the years.
Sadie Schoeffler Whitehurst: Absolutely! Ok well, thank you, Zac.
Zac Selden: Thank you.
Tad Britt: Thanks.
Read other Preservation Technology Podcast articles or learn more about the National Center for Preservation Technology and Training.