Minnesota Department of Transportation
Jump to main content.
Minnesota Department of Transportation

511 Travel Info

Snowy field
 

Mn/Model

Minnesota Statewide Archaeological Predictive Model

Contact Us   Mn/Model Home | Archaeology | Geomorphology | Geographic Information Systems (GIS) | Implementation

 

Quick Links

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Mn/Model Research Design
Appendix A: What is an archaeological predictive model

 

Mn/Model logo

An archaeological predictive model is essentially a map that indicates the relative likelihood of encountering archaeological finds in a specified region. Such maps are sometimes referred to as archaeological "sensitivity" maps because they indicate that some locations are more sensitive than others in terms of the presence of cultural resources. One can think of these predictive maps as showing three zones: for example, a high sensitivity zone where archaeological sites are most likely, a medium sensitivity zone where sites are less likely, and a low sensitivity zone where sites are unlikely.

 

 

These maps, if they are accurate, hold tremendous potential as planning tools. If new highway and other land-disturbing projects can be guided to regions of low archaeological sensitivity, then significantly fewer archaeological sites will tend to be encountered than otherwise. The result, in the long run, is reduced costs for archaeological survey, mitigation, and clearance.

 

It is the accuracy of these predictive models that determines their utility. Obviously, we would want most archaeological sites in a region to occur in the zone indicated as "high" sensitivity, very few sites in the zone marked as "low" sensitivity, and some intermediate number in the "medium" sensitivity region. This sort of performance can actually be examined and tested by comparing the model-produced maps to actual archaeological field survey results. By comparing model predictions against actual archaeological circumstances, it is possible to determine, with specifiable confidence, how accurately a model performs. It is, in fact, this very approach that gives us confidence in a model and allows us to use it as a predictive tool. Model testing, then, is an essential tool to demonstrate that a model has predictive utility and is not a figment of an archaeologist s imagination.

 

But how does a predictive model work? A simple example best illustrates the team's modeling strategy. Suppose we are interested in developing a model of modern-day camping locations. Being campers ourselves, we might surmise (theoretically) that good camp spots are located near water, close to a good fishing spot, on level ground, with a southeast-facing aspect (to capture the sun s warmth in the morning), near trees (as shelter from the wind), and perhaps on a raised setting to provide a good view. In fact, we might develop such a model through the use of interviews or a questionnaire of present-day campers. We could then take the indicated criteria and, through the computer-based mapping facilities provided by Geographical Information Systems (GIS), actually produce a map of the model that would indicate highly favorable, moderately favorable, and unfavorable campsite locations. Such a map would constitute a predictive model.

 

In the case of an archaeological predictive model, the basic modeling approach is similar, but with a few noteworthy differences. First, the native peoples and cultures whose sites are of interest to archaeologists generally existed far in the past, perhaps thousands of years ago. We therefore cannot use interviews and questionnaires to ascertain how past peoples located their camps, villages, and settlements. It is possible to do the next best thing, however, using the resources of archaeology.

 

Archaeologists can identify sites of past cultures and time periods using standard field-survey techniques. Taking samples of these sites, archaeologists can measure on maps or through satellite imagery relevant variables that can bear on their locations: distance to water, ground steepness, site aspect, landform shape, and a variety of geological, geomorphic, soil, hydrologic, and climatic factors. It is measurements made at sample archaeological sites that allow us to ascertain those criteria which influenced prehistoric site placement. For example, we might determine that prehistoric farming villages preferred a specific soil type, certain river terraces, and south-sloping land near the confluence of two streams. It is through this process that the past can "speak" to us and give us the equivalent of interview data. These techniques have been highly developed in archaeology and form a common regional research tool. Remote sensing scientists refer to this technique as "pattern recognition."

 

Once the pattern of relevant variables and measurements are known for a particular region and culture, it then remains to map it over the region to produce predictive maps. This is a process known in remote sensing as "pattern classification," and is conceptually simple to understand. Suppose, for example, that analysis produces a model that dictates that archaeological sites should occur 1) on a specific soils class, 2) on a level slope, 3) within 1500 meters (m) of water, 4) on an upper terrace ridge, which possesses 5) a southern or southeastern aspect (this is actually somewhat of a simplification). We could then go to a map and grid it into regular parcels of equal size, such as 50 x 50 meter units. In each grid unit each of the above five variables would be observed or measured on the map. If all five of the criteria are successfully met in the parcel, then that parcel is marked as "model specifies site," (perhaps by coloring it red); otherwise it is marked as "model specifies no site" (and not colored). This process is repeated grid unit-by-grid unit until an entire region is classified, with the outcome being a predictive model that indicates site-likely and site-unlikely regions.

 

Assessing Model Performance: What is Accuracy

The goal of the team's modeling effort is to produce computer-generated archaeological location models that map zones of low, medium, and high archaeological sensitivity. The twin issues of how accurate these maps will be and how accuracy itself is assessed are essential concepts that deserve comment.

 

Determining Model Accuracy

Conceptually, assessment of the accuracy of a model is a simple matter involving the comparison of what the model indicates about the archaeological situation, and what the actual circumstances are. In other words, the mapping of a model is compared against maps of known archaeological site locations and the percentage of known sites correctly "predicted" is computed. If, for example, there are 100 known archaeological sites in a region and 85 fall within the area mapped as medium and high archaeological sensitivity by the model, then the model has an accuracy rate of 85 percent. Ideally, the known archaeological site sample should be a random sample in order to trust that the computed rate of accuracy will reflect the model's performance on yet-to-be-discovered sites. This is one reason why new archaeological field survey based on random sampling designs is so important to the team proposal.

 

It should be clear that no model can work perfectly. If a model is constructed that, through testing, is shown to correctly "predict" 85 percent of known archaeological sites, then it will be true that approximately 15 percent of sites will not be correctly specified. This represents an amount of error that must be expected and allowed for.

 

The nature of this error is predictable, however, and stems from a fundamental methodological difficulty of archaeology. Archaeological location models tend to be biased toward the archaeological pattern of sites on or near the surface, and particularly toward sites of recent age. This is because a larger number of those sites have been found and recorded in the archaeological database. A portion of the team effort will be directed toward reconstruction of paleoenvironments and a consideration of geologically-buried sites to help mitigate this factor. Large-scale vegetative and climatic patterns in the past will be constructed using a climatic modeling tool recently developed by Dr. Reid Bryson of the University of Wisconsin-Madison. The tool links radiocarbon-dated pollen records to reconstructed precipitation and temperature for the state for the past 12,000 years.

 

Accuracy Versus Precision

To understand model performance one must also be familiar with the concept of precision. Good archaeological models must be accurate and precise at the same time. It would be easy to construct a highly accurate model, for example, by having it indicate archaeological potential at every location in a region (that is, every 50 x 50 meter parcel). In fact, such a model would be 100 percent accurate because it could not fail to predict every archaeological site that might occur in a region! This sort of model would, of course, be useless because it lacks any precision.

 

What is needed in practice is a model that maintains high accuracy when mapped, such as the ability to correctly indicate 85 percent of archaeological sites, but which at the same time maps substantially less than 85 percent of the landscape to medium or high archaeological sensitivity zones. A model that predicts archaeological sites with 85 percent accuracy in a mapped region that encompasses 85 percent of a region s area represents no real gain (one might as well throw darts at a map to predict archaeological site occurrence). On the other hand, a model that correctly indicates sites with 85 percent accuracy, but which "covers" only 40 percent of the landscape when mapped, represents a substantial improvement, in the form of a much more precise and useable model.

 

It should be noted that a model map can be displayed as a surface of continuously varying probabilities or can be divided into multiple zones, such as medium and high potential regions. This sometimes is of greater utility. Together, these zones might correctly specify approximately 85 percent of archaeological sites. Most of the sites will occur in the high sensitivity region, however, which will give a much more precise indication of where the bulk of the archaeological sites lay.

 

Role of Geographic Information Systems

Fifteen years ago, when this methodology was pioneered, this process was performed manually. Today, however, we now are able to automate the entire process through GIS computer technology, allowing all the necessary measurements and observations to be obtained for a model and the rapid production of regional maps of model predictions. This automation allows archaeological predictive models to be developed and mapped over very wide regions and at very high resolutions, such as 50 x 50 meter land parcels. Moreover, because the information is coded electronically, it is very easy to update and improve models in a timely manner.

 

 

 

 

MnModel Orange Bar Logo

Acknowledgements

Mn/Model was financed by the Minnesota Department of Transportation using funds set aside by the Federal Highway Administration's Intermodal Surface Transportation Efficiency Act.

 

Copyright Notice

The Mn/Model process and the predictive models it produced are copyrighted by the Minnesota Department of Transportation (MnDOT), 2000. They may not be used without MnDOT's Consent.