Scythian Mounds, Final Report

The final lab for the Scythian mounds project involved creating the final predictive results based on our previous work.

Using these surfaces, we created a feature class of random points (minimum 30 m spacing) throughout the study area and merged these points with the possible mounds identified from an aerial image.  This feature class was then expanded with fields and filled with the values from the reclassified surfaces (elevation, aspect, and slope).  These points were then subjected to an Ordinary Least Squares linear regression method based on these three surfaces (top map above).

The results of the OLS give the following results:

               Adjusted R-Squared:   0.760202

	Coefficient	Probability
Intercept	-1.770434	0.000000
Aspect	0.115368	0.000001
Slope	0.113834	0.000001
Elevation	0.630913	0.000000

The Adjusted R-Squared value of 0.760 suggests that the three surface variables can explain 76.0% of the sites in the predictive model.

Since the coefficient values are positive (with the exception of Intercept) and not very close to zero, it appears that each surface property is making a positive contributing to the model results (with our reclassified weightings).  The probability values < 0.05 also support this conclusion (near 100% confidence level for each coefficient).

We also subjected our OLS results to a spatial autocorrelation test, returning the following results:

z-score: 11.43385

p-value: 0

The very low (zero) p value indicates that it is very unlikely that the data is distributed randomly.  The high Z-score (in conjunction with the low p value) indicates that the data is normally distributed and spatially autocorrelated.

Based on a visual review of the OLS residuals and hot spot analysis, it does appear that many of the predicted areas are within valley bottoms.  It may be beneficial to add an additional variable that accounts for this spatial distribution, such as access to water in the form of rivers and streams.

It does appear that we have established that the variables in use are helpful for use in a predictive model. Per the lab, it would probably be a good idea to test these variables further with a larger point set and additional regression models (such as GWR).  Additionally the variables should be analyzed further to ensure the relationships are truly linear.  Adding additional variables (such as access to water) may also expose additional relationships.  Finally, any model should be subjected to some ground truthing to understand if it is truly useful.