 So first of all thanks everybody for for being here. I know it's early and a special thanks to the organizers for Allowing me to participate So I'm gonna be talking about absence makes the heart grow fonder So I'm gonna be talking about locating missing periods In our study area in specific landscape contexts And and how absence of evidence or negative evidence Inspired us towards using a geospatial approach and some of our preliminary results Yeah, okay, so a little bit about the research setting The project is run out of the Saqlossos research project parent project if you will And this is some of you may know the site But for those that don't we're in the on the northern flanks of the western tourist mountains in southwest Turkey In the city and Lake District, so the the tourist mountains run all along. We have pointer All along the coast of Turkey you have the tourist mountains which Separate the Mediterranean from the Central Anatolian Plateau and we're right on the northern cusp of that Saqlossos is an antique site from the first earliest dates are in the Hellenistic and it goes to the mid Byzantine But they've got generally three decades worth of archaeological research and ecological research in the territory Which was a great data set for us to draw from So the project recognized from the early stages the importance of Surveying to understand the context of the site and to survey for all periods In the in the early years in the 90s So in the in the early years in the 90s it was exclusively extensive surveys, so this was Driving the landscape and interviewing locals. What have you seen? Reviewing historic records and maps and these sorts of things But in the late 90s and early 2000s they began intensive survey programs So there were several separate separate projects over the years Focusing in different areas trying to get a good sample of the landscape most recently 2016 2017 myself and my co-author on this paper Ralph van den we did survey in the Highlands and the the result of this I don't know what you can see. I guess these they're quite small, but we've 200 and identified 265 sites throughout the territory So quite a large number of sites and I it's certainly only still a small fraction, but We find quite different archaeologies in different landscapes and For example in the Booter plane, I'm going to use the example of comparing the Booter plane Which if I go back Booter plane is here in green The survey and that there go Highlands is all the way on the western on the eastern side, excuse me So a great Degree of difference in elevation The Booter plane is your typical lowland fertile Alluvial setting that you expect to find a high degree a large number of finds In a place like Turkey But surprisingly to everyone we found quite a bit more in the Highlands and in the Lowlands and found periods which were entirely absent or very poorly represented in the Lowlands in the Highlands So we recovered more material two campaigns in the Highlands than in three in the Booter plane And this is the Booter plane is known for sites like Agilar and Courage Highway is famous Neolithic and early Catholic tell sites And when you compare the results from the two you can see that Having only focused on one or the other you would miss entire periods so you have the Hellistic for example, which is well represented in the Booter plane totally absent from the Highlands And the inverse is true for the late antique period and the difference in in Ottoman presence, for example, is also major and I don't have it on this Unfortunately, but we found quite a bit of Middle Paleolithic stuff also in the Highlands Which was known in the in the territory at all before our survey in 2016 So you can see this there that there are several blank areas in one landscape or the other Which you'd be entirely unaware of had to not sample the entire landscape or all separate landscape units So some possible exclamation I mean there can be many explanations for this right But some of the primary ones that I'll focus on I think are the diversity of the landscape of the Sakalosos territory you can see The extreme difference in elevation so we have Across a very short distance you go from the lowest of about 200 meters mid 200 meters to 2600 Just within 30 kilometers So local vegetation hydrology andography have created all these separate little micro ecologies and individual valleys are quite different from each other You can look at some of the photographs and we have everything from steppic planes to Badlands and then bamboo grass and marshy areas and then this oral Mediterranean Mount scapes as well. So it's a really diverse territory and mean annual precipitation for example between Sakalosos and the Buddha plane There's more than it's twice as high precipitation is twice as high annually in Sakalosos and the Buddha plane 30 kilometers away and There's a difference in an annual mean temperature of five degrees centigrade So it's a very diverse landscape and you see quite a lot of variability I think this has a big impact on and what we see the other is of course Erosion human impact that have affected the visibility of remains all right What can we what's being obscured by the way the landscape has changed over time? Human activity on the hill slopes has caused really extensive erosion in many of these valleys Have trend they transported sediments several meters thick in places We know some of the neolithic layers by dating Course that we've that we've extracted that the neolithic labor layers in places are eight meters below the surface the current surface now So and also material of course is transported in this process as well. So obscuring certain periods in certain places, but Graveling with these these gaps in the record has sort of Inspired us to take a different approach at finding a way to target specific periods and types of sites To help you know fulfill our research goals in general And that approach that we that we decided to use utilize Which was designed by my second co-author Chris Carlton at all is known as Lamar and the The theory behind Lamar is that humans make land-use decisions based on by referencing mental archetypes which come from practice tradition social memory And that the archetypes are basically schematics Which to pick to them optimal locations for given activities whether it's clay sourcing or Flint napping or Building a homestead or campsite and that there's an inherent proximate bias in this right? It's relative to the space around you. So you're not thinking of the best campsite possible It's what is best in this area. Maybe a rubbish campsite 50 miles kilometers away But this is the best one available to you at the time. So this is an inherent proximate bias in the model accounts for that in terms of calculation The basis of these empirical frequency distributions So we start with the training data set the training data set is as the database of all the known sites their locations their period their size And a selection of predictor variables Which the researcher determines which it which we think are the important factors which are affecting settlement choice So we started with a basic set. I should say that this is our we this was our first run this summer I'm fresh from the field. So we have only preliminary results But what we this can be tweaked over time So the the variables that we chose to use were elevation slope aspect convexity proximity to drainage As landscape variables and then cultural variables for certain periods such as proximity to least-cost paths or proximity to urban centers proximity to Roman roads, etc and Then these variables the values for these variables are extracted using the GIS Which from which we derive the multivariate distribution traits for each known site The output is a map a roster image where which each cell with each pixel Is compared to these multivariate distributions for each each cell is compared to the multivariate distributions for each known site so you're making compare you're making comparisons and The result is probabilities of a cohesion Which are combined and weighted and they give us a number which represents archaeological potential so here's some of the Surfaces the predictive surfaces that we generated to show you an example of how they differ the top left, sorry the top left is is the like prehistoric predictive surface and The then the top right is a Hellenistic bottom left We have a late antique and the bottom right is autumn and so you see there's quite a bit of variability You see that the bottom left and top right are much more constricted And this is because it's it's an interpolative model and we use a factor like proximity to urban centers So it's bound by the extent of urban centers But if you were to zoom in on one of these images Very closely in a GIS you see that although some areas look entirely white or entirely dark when you get up close There's quite a range of variability Even from one cell to the next Our methodology the way we went into the field where we worked with the models or were to select cells with the goal of achieving a Representative sample of each from that class so in addition to Seven models one for each of major period that we have the ones that showed you We also produce an aggregate Model which combined all of them to give us an aggregate score Which was zero through 35? These are the Lamar classes you're through 35 35 representing the highest archaeological potential Zero obviously the lowest and for each class each individual period. It's a one through five So we had the task which was quite a challenge to To try to get at least one from each Lamar class of the 35 And then if we got one then we try to get to get to try to get three And at the same time to try to get zero through five represented for each period So it was it was I spent hours in the lab in the evening after surveying all day trying to plan out the Survey for the next day in the future. I think we'd like to try to find a way to to create an algorithm or something Which we can automate that But we tried our best and in the end We were able to to survey 93 cells Which I think was quite good in the time that we had The cells were selected to from a diverse range of landscapes We tried to get lows mids and highs from each different landscape unit that we have and Each cell was then Intensively surveyed I have an image here. This is our survey methodology and So we used a total collection strategy and each sherd collected was dated by our on-site specialists to within the Lamar Ranges of time So some of our preliminary results, like I said, this is the beginning But the bar graph here shows the ratio of positive plots to the total number of plots for each of map class For each time period the height of each part is effectively the empirical probability of finding something giving up give a Lamar class Since the the the ratios are higher for high Lamar classes We can conclude that the predictive model is generally working across all time periods though some periods appear to be better than others It should be noted that despite our best efforts obviously and I described that we've limited time and the complicated Survey planning process meant that we couldn't get complete coverage For each class field. So we don't even sampling across all the map classes For each time period but the results Are also complemented by the fact that some periods are are they're better represented in the archaeological record than others So some periods are just harder to find in general which weighs the model down For example, the iron age is quite difficult for us to find at all So it's bringing these totals down which are the aggregate totals You can see here also in these In these plots which show the same basic trend these are Showing the the number of fines. Okay, so this is not just present absence But numbers of fines and they're showing also generally that aside from some outliers that it's working fairly well So two preliminary statistic models have been created to explore the survey results Luget regression model which involves a binary outcome. This is the presence absence It estimates that the effect given the predictive variables has on the odds of a positive outcome And the poisson regression model which involves the count-based outcome And it estimates the impact given by predictor variables on that count So far analysis the lemmat classes were used as a predictor variable The question we're trying to answer is do higher lemmat classes mean improved odds of finding something a presence in legit Model or finding more stuff the poisson model for this initial analysis We didn't treat each period separately instead. We looked at it in aggregate for both models we use a generalized linear regression and AIC To select the models we compared the AICs in each case to corresponding models that let the lemmat class a variable out of the equation and so models that used In only an intercept the the mean to predict the outcome and compare that to the AICs for the models with using the lemmat variables And we found that the lemmat variable results were much better Yeah, yeah Okay. Yeah, so Here you can see in the end In any one unit increase in the map for us in our territory showed if you go from a one to two You get a 27 percent higher chance of finding something presence absence a four unit increase would give you An improved odds by a hundred and sixty one percent that you find something One unit increase in the map class Also corresponds to a 50 percent higher increase in the number of finds So going into the field with this in hand I think would give us a much better results in terms of in a large territory finding missing periods Future analyses we've got many plans how we want to go forward with it things we can do I'll try to wind up here, but I Including new new survey data looking at different variables and the effect each variable had and Trying to tweak a little bit how we can get the model to produce the exact results which we had This is an image of our aggregate models. You can see what the aggregate model looked like for the territory Just to give you an idea and and that is That is it