GIS 5990 Final Project: Predictive Model of Neolithic Sites in Southeast Turkey

Figure 1: Overview of study area.

The aim of this project was to create a predictive model of Neolithic archaeological sites in southeastern Turkey. This area was chosen due to the location of Gobekli Tepe in the region. This site has become well known as one of the earliest uncovered monumental structures with some researchers (such as excavator of the site Klaus Schmidt) hypothesizing that mobile hunter gatherers constructed the site. The impetus for its construction may have been a change in ideology that also created fertile ground for the transition to sedentary, agricultural lifeways.

Figure 2: Site in the study area.

A successful predictive model would assist in explaining site location and help researchers in discovering new sites. Site data were acquired from The Archaeological Settlements of Turkey GIS (http://tayproject.org/giseng.html). Neolithic site data were manually entered into a new shapefile in ArcMap. This data is the weakest link in the analysis due to the potential for spatial inaccuracy.

The digital elevation model (DEM) seen in figures 1 and 2 were created by using the Mosaic tool on fifteen ASTER images acquired from the USGS Earth Explorer website (http://earthexplorer.usgs.gov/). With the image properly projected as a DEM, slope and aspect data could be generated in addition to elevation data.

Figure 3: Elevation data reclassified.

Based on the location location of Neolithic sites at lower elevations, the Reclassify tool was used to classify the elevation data into four equal categories. Category 4, representing the category with the highest probability of sites, was from 0-800 meters. In descending probability, category 3 was 800-1600 meters, category 2 was 1600-2400 meters, and category 1 was 2400-3200 meters.

Figure 4: Original slope data.

Figure 5: Original aspect data.

Figure 6: Slope data reclassified.

Figure 7: Aspect data reclassified.

Slope and aspect data were similarly classified into four categories of high to low site probability. The highest probability category for slope was defined as 0-5%. In descending order of probability, the remaining three categories were 5-15%, 15-25%, and 25-75%.

An aspect of 0 (or flat) was rated as the highest probability, with southern-facing aspects being the next probability category. East and west facing aspects were the next category, with northern facing aspects being of the lowest probability.

Figure 8: Water buffer of 500 meters.

Data on rivers and lakes, acquired from Eurostat (http://epp.eurostat.ec.europa.eu), were buffered to 500 meters. A raster was then created by using the Feature to Raster tool with probability defined as binary categories. Areas within the 500 meter buffer were considered high probability with all else being low probability.

Figure 9: Weighted Overlay.

After the data on the four variables were suitably processed, the Weighted Overlay tool could be used. Elevation and water source distance were each weighted 35% while aspect and slope were each weighted 15%. The result can be seen above.

Figure 10: Ordinary Least Squares analysis.

In order to test the weighted overlay predictive model, an ordinary least squares analysis would be conducted. First, the Create Random Points tool was used to created over 100 points to function as locations without sites. This was merged with the original site shapefile. The Extract Multi Values to Points tool was used to incorporate data from the variable rasters into the merged point shapefile. The Ordinary Least Squares tool can then be run.

Unfortunately, the statistics indicate the model is not successful in explaining site locations. Only elevation returned as a significant variable. There are several explanations for why the other variables were not significant. First, as noted above, site location accuracy is not assured based on the data source used. A more robust model would incorporate verified site locations. Additionally, the reclassifications of slope, aspect, and water source distance may have been incorrect. Other classifications may function better. It is also possible that slope and aspect do not in fact help explain site location.

We might also attempt to incorporate paleoenvironment data, including on water sources. Using modern hydrography as a proxy for ancient water sources may not be justified.

Another potential variable to include in future models is viewshed and site visibility. Figure 11 shows an experiment with the Observation Points tool depicting raster cells visible from five well-known sights in the study area. Gobekli Tepe in particular has a wide view to the south. There is potential here for future model attempts.

Figure 11: Observation points for five site in study area.