Introduction
Location Map
Base Map
Database Schema
Conventions
GIS Analyses
Flowchart
GIS Concepts
Results
Conclusion
References

Conclusion

Conclusion

The results from this experiment are largely inconclusive. Nevertheless, we can draw two broad conclusions:

1. Human observers are generally not reliable at drawing boundaries around complejos in GIS, and experience does not appear to have any significant impact on this. This means any map drawn by hand based on a LiDAR generated DEM of an archaeological site must be treated as hypothetical at best, and stresses the need to develop more rigorous methods of mapping.


2. The Multiresolution Segmentation algorithm from eCognition is far more precise than a human observer, and is capable of detecting patterns at the specified spatial scale with a high level of consistency. However, whether or not it is more accurate is still inconclusive. The algorithm is only capable of analyzing the data that it is provided in the raster, and since road networks and plazas were not clearly demarcated in the rasters the segmentation did not recognize them as boundaries.

Under the paremeters of this study, we cannot reach a definitive conclusion about the effectiveness of the MRS algorithm, but we can suggest a few steps that may improve the algorithm's effectiveness.

First, additional kinds of information may be coded in the raster images. All rasters analyzed by the algorithm in this study were exported as RGB rasters from GIS. This has the consequence of combining the pixel values for each of the component rasters into a single value representing color. A better method would be to export the relevant rasters as grayscale rasters and combine them in ENVI. This would allow us to code particular kinds of information in particular bands, and other bands could be used to record additional information. For example, if we had a digitized map of road systems at the site, we could code these as one band of the raster and the algorithm would subsequently include roads as part of the parameters for determining complejo boundaries. This would likely increase the effectivness of the algorithm substantially.

Second, eCognition allows us to keep segments from one iteration in subsequent iterations. So, for example, we could select all segments from the output that have greater than 80% overlap with their corresponding polygons in the original complejo map. Then we could adjust the algorithm parameters and run it again, and the algorithm would keep the complejos we've selected and run the subsequent iteration on the remaining parts of the image only. This would allow us to recognize patterns at multiple scales. We chose not to do this for the purposes of this experiment, since the goal was to compare the algorithm against a human observer. This method would effectively reintroduce the subjective bias of the human observer in determining complejo boundaries, but this may prove more effective than either the human observer or the algorithm alone.

In the end, although the results of this study are inconclusive, they are nevertheless useful. They suggest that there is some validity to this line of inquiry, and provide us with useful suggestions for how to in increase the effectiveness of the use of the MRS algorithm for mapping complejos.