Special Topics: Module 11 - Final Project (Modeling Cave Sites)

     This final project is the culmination of the work we have carried out in the name of GIS for archaeology. We have modeled Scythian landscapes, areas potentially containing shipwrecks, and site probability in an archaeological district. For my final assignment I wanted to model potential areas where ritual cave sites could be located in Greece. As it turns out, my model fails to reject a null hypothesis of random site distribution. My data is not clustered or spatially autocorrelated. This is likely due to a small known sample size and a simplistic model based upon only three explanatory variables (slope, aspect, and elevation). Each map below represents progress in the development of a predictive model.
     Data for this assignment is freely available and was compiled from several sources. Sacred cave data was partially derived from The American School of Classical Studies at Athens. Additional cave sites were compiled from Wikipedia and a caving enthusiast's website. Digital elevation data was derived from ASTER tiles downloaded using USGS's Earth Explorer. (When downloading lots of data be sure to use their handy bulk downloading tool).

Study Area
    The extent of the study area was determined by the cave data. An arbitrary polygon was drawn around a portion of mainland Greece and Crete. The cave data represent multiple periods of ritual use (from Neolithic to Roman). The point locations are also generally approximate and unverified. I used the Mosaic to New Raster tool to stitch all 30+ ASTER tiles together into one raster. I then clipped the raster to the study area polygon. The cave shapefile was a product of the ASCSA data and the other site data. I used an Excel spreadsheet to add XY data and merge it with the ASCA data. Voila! More data (relatively speaking, my sample size went from 10 to 21). 

Map 1. An overview of the study area with a DEM
overlaid atop an Esri imagery base map.

Map 2. An overview of the study area with sacred
cave sites noted. 

Primary Coverages
     Before secondary coverages can be generated, it is necessary to examine patterns in existing data. This helps in deciding weights to assign when data are reclassified. Here we see a slope, aspect, and contour coverage of the study area. Each are derived using the DEM and their respective Spatial Analyst tool (Slope, Aspect, Contour). There were not any obvious patterns in the data particularly concerning aspect. The caves were almost equally distributed between northerly and southerly aspects. Most caves were located between 0 and 600 meters in elevation. They were also predominately located in areas with a slope between 10 and 20 degrees. There were outliers though. The Cave of Zeus was located at a much higher elevation and slope than any of the other caves (only the highest and slopiest for old Z-man). This information was taken into consideration when developing weight classes for data ranges.
   

Map 3. A slope surface derived from DEM data.

Map 4. An aspect surface generated from a DEM.

Map 5. A map showing contour lines in 300 meter intervals.
Cave sites are shown to demonstrate the process of
noting characteristics of cave sites.

Secondary Coverages
     Once the data have been examined and weights decided upon they are used to develop secondary coverages. This is done using the Reclassify tool in ArcMap. Here we see the results of that process. Each map shows probability represented by three classes. Areas possessing similar characteristics to known sites are symbolized in red denoting a higher probability of finding sites, moderate probability areas are in yellow, and low probability areas are in green. These coverages will then be weighted to develop a predictive model.

Map 7. Slope reclassified. 

Map 6. Aspect reclassified.

Map 8. Elevation reclassified. 

Predictive Model
     These secondary coverages are now ready for use in deriving a predictive model. I generated multiple predictive model outputs using different weights assigned to each secondary coverage using the Weighted Overlay tool. Each model emphasizes one coverage over the others. The exception is Model D which weights each coverage equally. This comparative exercise shows what happens when weights of explanatory variables are altered. When using predictive models to inform survey and archaeological investigations it is important to generate a solid predictive model. These models have common areas of high probability but differences could mean a lot when time and money for a project are of the essence.

Map 9. The predictive models. Each one of these models are the result of assigning different weights to the secondary coverages. Weights assigned are noted beneath each data frame. High probability areas are symbolized in red.

Analyses
     Predictive models cannot be taken at face value. They must be statistically examined. Turns out, when I probed my data it was found wanton (but that doesn't mean it isn't informative). The data is random and not spatially autocorrelated. It overpredicts sites in low elevation areas especially on the smaller islands and Crete. I performed an Ordinary Least Squares regression (in order to perform these analyses I had to generate random non-sites as a proxy for known non-sites in the study area), spatial autocorrelation, and a hot spot analysis. The results of these tests are seen in the final map product for this assignment.

Map 10. OLS, hot spot, and spatial autocorrelation results. 

Special Topics: Module 9 - Biscayne Shipwrecks (Part 2)

     All that preparation last week set the stage for analysis this week. So we'll dive right in. First, I defined a 300 meter radius surrounding the shipwrecks to examine the benthic environment surrounding each wreck. Then, reclassified rasters were generated for the benthic and bathymetric data. These were used in conjunction to create a simplified weighted overlay/predictive model.
For this analysis, there was not enough data for a linear regression or a statistically meaningful analysis of the shipwrecks (we were working with n = 5) but we had enough data to guide survey efforts.
     As in previous weeks regarding predicitve models, essential tools were the Feature to Raster tool,  Reclassify, and the Weighted overlay tool. As well as the workhorses of geoprocessing: the Buffer and the Clip tool.
     The three map deliverables below each show a different aspect of the analysis. Map 1 is a summary of the benthic characteristics within the study area (300m radius) surrounding each shipwreck. Map 2 shows a comparison of the reclassified data used to derive the predictive model. Map 3 is the predictive model. In Map 2, areas in red denote high probability while those in dark green represent the lowest probability of finding a shipwreck site. The areas in red in Maps 3 are thse areas where the model predicted areas with a high likelihood of containing a shipwreck. That is those areas most resemble the conditions surrounding the known shipwreck sites.

Map 1 shows the benthic environment surrounding the study area of each site.
It also contains an overview of the park.

Map 2 shows the reclassified raster data used in the predictive model.

Map 3 is the end product of our analyses. This is the predictive model
generated by using reclassified benthic and bathymetric data.

Special Topics: Module 8 - Biscayne Shipwrecks (Part 1)

     Time to sink in (bad joke) to some shipwrecks. That's right. These ships went down near Biscayne Bay, Florida. They don't make jokes about you for nothing Florida. Way to s(t)ink, I kid. I love my home state. This assignment made me feel a twinge of home sickness. Luckily I will be heading that way soon and maybe snorkel this maritime heritage trail.
     This week I explored two websites looking for data to use for this marine archaeology project. The first website provided the historical chart while the latter is where I acquired DEM data for Biscayne National Park. Point data for the heritage trail was provided to us care of UWF as was the benthic data.  Finding the DEM was a bit of a chore as the data I thought I was accessing led to another website/tab several times before I actually reached the data. And, often that data wasn't what I was looking for. Perseverance is the order of the day. Once I downloaded the historical chart I georeferenced it. The DEM/bathymetric data was clipped to a digitized boundary of the park. The data is now prepped for use in the coming weeks.
     The final map deliverable is seen below. It shows the benthic data, historic chart, and DEM data all clipped to the park boundary. The shipwreck locations display in each data frame giving them context in different visualizations within the same space. A short description is also included on the map that briefly discusses the map content.

A map deliverable displaying data for Biscayne National Park, Florida
 prepared for analysis in the coming week. 

Special Topics: Module 7 - Modeling Scythian Landscapes (Part 3)

     This wraps up our discussion of Scythian landscapes. We prepped data, generated secondary coverages, and this final week we used Ordinary Least Squares (OLS) regression to explore our data.
The first task was to generate non-sites (though this is not an ideal method as it is not ground truthed data). ArcGIS has a nifty tool that generates random points across your data. The Create Random Points (a Data Management tool) accomplishes that end. Once the points were generated they were merged with the Tuekta mounds identified last week. The attribute table was then populated with the reclassified slope, aspect, and elevation numbers from the secondary coverages.
     With the attribute table populated, I ran an OLS linear regression using these sites and 'non-sites.' Turns out the model, as it stands, explains about 74% of variability in the data. (Elevation had the highest coefficient of the three variables.) We also checked the results of the OLS regression against the results of a Spatial Autocorrelation and a Hot Spot Analysis. The results are summarized in the map deliverable. Two data frames display the OLS and Hotspot analyses over a DEM of the study area. The results (a low p-value and a high Z-score confirm the results of the regression - the data is clustered and non-random) of the Spatial Autocorrelation are shown at the bottom of the map. Adding additional variables could improve the predictability of the model, however, these variables are enough to parse the statistics and learn a bit about probing landscape level archaeological data.

Map deliverable showing the results of both the OLS and
hotspot analyses overlaid atop the study area DEM.
The Spatial Autocorrelation results are also shown for reference.

Special Topics: Module 6 - Modeling Scythian Landscapes (Part 2)

     The week prior focused on data preparation for this lab. This week secondary coverages were generated for slope, aspect, and elevation. In addition, contour lines were also derived from the DEM. The georeferenced image of the mounds was used to generate a shapefile of the point locations of the mounds. This was great practice for identifying archaeological features in the landscape and also to discern any patterns these features may or may not possess.  The general trend seems to be that the mounds are found in areas with low elevation, facing in southerly directions (SE, S, SW), and they lie at slopes between 0 and 10 degrees. Next week, the lab will statistically explore the data and we will 'publish' it.
     The map deliverable below shows the several coverages that were generated along with a shapefile showing the point locations of some of the Tuekta mounds. The coverages were generated just as was done in preparation for the Tangle Lakes project weeks prior.

Special Topics: Module 5 - Modeling Scythian Landscapes (Part 1)

     This module is all about data preparation (for labs in the coming weeks) and starts our foray into landscape archaeology. We walked through gathering data for international landscapes from the USGS Earth Explorer. In particular, locating DEM data for an area in the Altai Mountains of Russia that contains known Scythian mound burials (also known as kurgans). For this exercise we used ASTER derived data. The Mosaic to New Raster tool was used to stitch together four raster datasets and from there we digitized a boundary and clipped the larger DEM down to something more manageable.
     The map deliverable shows the subset of data examined by way of a DEM symbolized with the elevation color ramp. It also highlights the area where the mounds are located with a georeferenced JPG.

This week's deliverable highlighting a region in the
Altai Mountains, Russia that contains burial mounds.

Special Topics: Module 4 - Predictive Modeling

     Predictive models are another tool in the archaeologists toolbox and this week we ran a simple predictive model for an archaeological district in Alaska. Predictive models can take many factors into account but for the purposes of this lab we used four components/generated surfaces: slope, aspect, elevation, and proximity to water. The Spatial Analyst extension is key in implementing a predictive model in ArcMap, particularly the Reclass, Surface, and Overlay toolsets (for the Weighted Overlay tool).
     While predictive models are great for informing survey decisions and streamlining cultural management processes they're aren't a panacea. They can in no way serve as substitute to boots on the ground, they can merely supplement and inform survey and project planning. There is no model, no matter how thorough, that can say that there are no archaeological materials in an area. These models represent probability and likelihood of the presence of archaeological remains.
     I am currently at the Great Basin Archaeological Conference and one of the presentations I was able to see discussed a predictive model for an area of National Forest that I work in. The model had a greater than 90% ability to predict the known site locations (it was thoroughly cross validated and I have spent the summer collecting data in some of the unsurveyed areas...it will be neat to see if the model predicted a comparable amount of density in that region). That being said, it's those sites that appear in the unlikely areas that become of great interest (in their deviation from a pattern).
     Yes, predictive models do not tell the whole story but they go a long way in helping, in my instance a National Forest, determine how best to allot limited resources (time, money, seasonal employees).
     Below is my map deliverable showing the areas that are likely to contain sites in green. That means their combination of slope, aspect, elevation, and proximity to water are ideal.

Map deliverable displaying the results of a predictive model
run using slope, aspect, elevation, and proximity to water.