Mn/Model Final Report Chapter 9: Project Applications of Mn/Model

Quick Links

Chapters

Appendices

Chapter 9

Project Applications of Mn/Model

By Craig M. Johnson and Elizabeth Hobbs

Statewide Survey Impelmentation Model Map

Chapter 9 Table of Contents
9.1 Introduction
9.2 Comparison of Mn/Model vs. Intuitive Model
9.3 Comparison of Two Generations of Models
9.4 The Phase 3 Model
References

9.1 INTRODUCTION

This section presents the results of an application of Mn/Model to an archaeological reconnaissance project in Stearns County. It involves the production of site model maps combined with elevation and surface hydrology. These maps are a useful adjunct to the small-scale subsection maps, providing archaeologists useful information at a scale that can be used in field investigations. The first application compares the Phase 1 site model and an intuitive model of site location based on past experience of an archaeologist. The second series of maps presents the Phase 2 site model and compares it to the Phase 1 model of the project area. The final discussion compares a third set of maps, depicting the Phase 3 site model, to the Phase 2 model.

Return to Top

9.2 COMPARISON OF MN/MODEL PHASE 1 MODEL VS. INTUITIVE MODEL

The following is a case study comparing a Mn/Model Phase 1 model to an intuitive model of a professional archaeologist. The comparison was carried out in east central Stearns County as part of the T.H. 23 project (Johnson and Maki 1996:5.6-5.14). Its purpose was to impose an element of ‘ground truth" on the preliminary results of Mn/Model, a necessary element in any application and refinement of this predictive model.

The T.H. 23 project falls within the Central Lakes Deciduous archaeological region (Anfinson 1990). A small portion of the Phase 1 predictive model developed for this region, encompassing the locality around the T.H. 23 project area, is the focus of the following discussion. Depictions of this model appear in Figures 9.1A, and 9.1B. These maps include preliminary site probability areas (high, medium, low) developed by Mn/Model Phase 1 and the "intuitive" probability areas, again divided into high, medium, and low classes (Johnson and Maki 1996:Figures 5.5A-5.5B). These maps also depict archaeological sites, lakes, rivers, streams, and topography in the form of elevation contour lines. Topography and surface hydrology are the two major variables that enter into both Mn/Model and intuitive models of site location.

The Mn/Model Phase 1 results (FMOD1) were developed using data from Stearns, Wright, Anoka, Isanti, Washington, and Chisago counties, and from parts of Carver, Douglas, Crow Wing, Clay, Kanabec, Dakota, Pine, and Cass counties. Keep in mind that this model had a gain statistic of only 0.22, so is not expected to perform much better than by chance. The intuitive model is mapped only within 200-foot corridors along the existing (EX), southeast 2 (SE2), southeast 3 (SE3), and railroad (RR) alignments. A 200-foot corridor was chosen as a rather arbitrary figure since the actual corridor width varies widely through the project area. It generally represents the widest construction limits and is also easy to represent on a map.

There are several ways a comparison between the Phase 1 model and the intuitive model can be approached. One is to examine the maps in Figures 9.1A and B for corresponding and conflicting site sensitivity values. Starting with the Phase 1 model (Figure 9.1A) and moving from west to east, there is good agreement between the models where the Sauk River flows into the chain of mostly artificial lakes south of Richmond. From there, the Phase 1 model depicts those areas of low potential in the intuitive model as medium to high. The two models generally agree along the RR alignment. One half mile east of where the RR alignment rejoins the EX corridor, there are considerable differences. This is an area of low relief and drained wetlands that are generally considered to be amorphous landforms not suitable for human occupation in the intuitive model, but considered to be medium to high potential in the Phase 1 model.

There is generally a good correspondence between the two models where the Sauk River passes under the T.H. 23 EX alignment. Over the next several miles, there is a major difference between the two models. The intuitive model ranks much of a terrace of the Sauk River as medium potential, whereas The Phase 1 model depicts it as low probability (Figures 9.1B). The same kind of discrepancy is found further east, on the Sauk River terrace along the EX alignment north of Rockville. Finally, a similar difference between the two models can be found at the northern end of the project area along the second terrace of the Sauk River.

One explanation of these discrepancies is that the intuitive model is based on where the Sauk River could have flowed in the distant past (e.g., about 3000 years ago or more), whereas hydrologic data for Mn/Model depict only the current course of the river. Also, the geomorphic data providing locations of river terraces were at a resolution too coarse to map this particular terrace. Even the 1:100,000 DNR landform coverage distinguishes only one terrace, north of the river and well away from the road alignment. In addition, the SHPO archaeological database is largely made up of sites that are assigned, or can be reasonably assumed to be assigned, to the Woodland period (400/200 B.C. - A.D. 1650). This is apparent in the sites west of Rockville that were recently found during a pipeline survey (Murray et al. 1995; Breakey and Dobbs 1995). The majority of these can be assigned to the Woodland tradition. All are near the Sauk River. The dominance of more recent sites in the archeological database and, because of survey biases made explicit in the survey probability models (Chapter 8), their proximity to present-day water bodies causes Mn/Model to do a poor job replicating intuitive models for locating older sites. Thus, as distance from the Sauk River increases, site probability in the Phase 1 model decreases.

Another discrepancy between the models is between Cold Spring and Rockville, where the EX alignment butts up against a series of north-facing slopes. The one level area is where T.H. 23 and the railroad have been excavated into hills. Because these slopes are generally uninhabitable, they received a low intuitive value. The Phase 1 model classifies these steep slopes as high probability. This can also be observed on other portions of the maps, particularly northeast of Richmond. Because smooth terrain is associated with poorly drained soils, archaeological sites are often statistically associated with rugged terrain. This is a situation where the common sense of the archaeologist is expected to override the model. If terrain is too rugged to support habitation or the preservation of archaeological sites, it will not be surveyed no matter what its probability class.

Progressing east, there is modest agreement between the two models up to Rockville. A portion of the SE2 and SE3 alignments southeast of Rockville receives a low intuitive value because of its distance from the river and a generally high value from the Phase 1 model. A perennial stream entering the Sauk River from the south and the bluffs that define the limits of its terrace to the east account for the Phase 1 model's interpretation of this area.

There is closer agreement between the two models at the northern ends of SE2 and SE3 as they enter the Sauk River valley. The two models generally correspond in the large low probability area south of Rockville along the SE2 and SE3 alignments. It can be noted that the northern border of the Phase 1 model's low probability area north of the Sauk (Figure 9.1B) is relatively straight. This unusual pattern corresponds to relief features in what appears to be an abandoned stream channel, perhaps even a former channel for the Sauk River.

Finally, there is a major difference between the two models in the uplands at the end of the project where the intuitive model classifies the Phase 1 model's high and medium probability areas as low. A portion of this area is characterized by wetlands separated by low, large, amorphous rises. These areas are thought to have little attraction for precontact occupation in the intuitive model. Since this Phase 1 model was built using data from a very large region, it is not environmentally homogeneous. Distance and direction to water or wetlands are among the model variables. Parts of the Central Lakes Decidious region, like the Anoka Sand Plain, are characterized by lakes and numerous wetlands interspersed among flat to rolling topography. Sites from these areas may have caused the model to place more emphasis on wetlands than would be appropriate for this study area. Models based on smaller, more environmentally homogeneous subregions may resolve this problem.

Another way to evaluate the relationship between the intuitive model and Mn/Model is to construct a table of intersecting values, as in Table 9.1. This table lists the land area in hectares (ha) classified in each probability area under each model. One hectare is 10,000 m2 or 2.471 acres. In considering this comparison, keep in mind that the Phase 1 model was constrained by the rule that approximately 33 percent of the land area be in each of the high, medium, and low probability classes. The intuitive model had no such restraint. It maps only nine percent of the land as high, but 38 percent as medium probability. However, if the models are similar, one might expect that the majority of areas mapped as high and medium probability in the intuitive model might also be included in the larger area of high and medium probability mapped in the Phase 1 model.

Of the total of 19.66 hectares placed in the intuitive high probability area, the Phase 1 model classified 9.10 acres as high and 6.20 hectares as medium probability. This indicates that the Phase 1 model agrees with the intuitive model's high probability class 46 percent of the time and considers the intuitive model's high probability class to be medium probability 32 percent of the time. Large discrepancies (high probability in the intuitive model and low probability in the Phase 1 model) occur 22 percent of the time. Areas that the intuitive model considers medium probability are classified by the Phase 1 model as either high or medium probability 65 percent of the time. Thus, agreement between the models is relatively weak.

A numeric value can be assigned to this relationship using a correlation coefficient called tau b (Anderson and Zelditch 1975). This is a measure of association between two ordinal variables and ranges from -1 (perfect inverse or negative relationship) to +1 (perfect direct or positive association). A calculation of tau b from Table 9.1 yields a value of 0.001 or no correspondence, one way or the other, between the two models. This does not mean that the two models disagree, however, since the value of tau b is not a negative one. However, this may not be a valid analysis since the proportions of the areas classified into each probability class differ so greatly. Phase 2 models for this region classify no more than 24 percent of the landscape as high and medium probability and may provide a more valid point of comparison.

An additional test is to examine the placement of sites. All sites depicted in Figures 9.1 A-B are located in Phase 1 model high and medium areas, except 21SNg. This site is described as a precontact artifact scatter. Tradition, context, and function are not known. It is situated on an island in the middle of an artificial lake, which would not have been part of the Mn/Model database. This is a good example of a site that is not related to its current environmental situation. Until a way is found to better model past environments, such sites may be poorly predicted by Mn/Model.

Of the three sites located during the T.H. 23 survey, but not depicted in Figure 9.1B (21SN124 - 21SN126), all are situated in Phase 1 model high to medium probability areas. This indicates a close correspondence between high and medium probability areas and site location when the Phase 1 model is applied to a specific locality.

Table 9.1. Cross-Tabulation of the Number of Hectares in the Three Site Probability Areas Based on Mn/Model and the Intuitive Model

Mn/Model Phase 1
		High	Medium	Low	Total
Intuitive Model	High	9.10	6.20	4.36	19.66
	Medium	18.87	33.97	28.94	81.78
	Low	44.66	29.09	39.88	113.63
	Total	72.63	69.26	73.18	215.07

Several conclusions can be drawn from this limited application of Phase 1 results. First, because of the way Mn/Model evenly divided the landscape into the three probability areas in Phase 1, there is a large portion of land that falls into the high and medium probability areas that most archaeologists, and later models, would assign to a lower category.

Second, some portions of the landscape, particularly those along bluff edges having steep slopes, could be eliminated from the high and medium probability zones. Extensive areas with the steepest slopes were eliminated from the models in Phase 3. However, because of the low level of accuracy associated with site centroids, many archaeological sites that actually occur on flat land at the tops or bottoms of bluffs are mapped on the bluff slopes. Until site locational data are improved, Mn/Model will continue to find associations with steep terrain that are spurious. Archaeologists must use common sense when interpreting the models. Other portions of the landscape have seen extensive commercial, residential, and industrial development that would effectively create low site potential zones in otherwise high and medium probability areas. These may be considered for exclusion from the Phase 4 models, now that relatively recent high resolution digital land use data are available. Until then, archaeologists must rely on their common sense when determining what areas to survey.

Third, Mn/Model relies on the SHPO database, which is dominated by confirmed or suspected Woodland tradition components, and modern environmental data. This creates a model that emphasizes recently developed landforms and underestimates those areas containing older, more rare, and in some cases more important, sites. Several methods are being considered for Phase 4 that will incorporate more information about earlier environments. However, with the dominance of Woodland sites in the archaeological database, features from past environments may not have much predictive power. If the actual number of Woodland sites greatly outnumbers those of earlier periods, Phase 3 models may represent a relatively accurate picture of site probability.

Finally, the models may perform better locally if they are developed within environmentally homogeneous regions. This problem was addressed in Phase 3.

Return to Top

9.3 COMPARISON OF TWO GENERATIONS OF MODELS

The Phase 2 basic model for the Central Lakes Deciduous South model (B2101), as applied to the T.H. 23 area, is depicted in Figures 9.2A and 9.2B. There is more agreement between the Phase 2 model and the intuitive one, particularly along the older terraces of the Sauk River, than the earlier Phase 1 model. This occurs in the medium probability areas of both the intuitive and Phase 2 models.

A careful comparison between Phase 1 (Figures 9.1A, 9.1B) and Phase 2 (Figures 9.2A, 9.2B) models reveals some dramatic differences. First, there is a substantial reduction in the high probability areas from Phase 1 to Phase 2. This was the direct result of changing the rules for model classification (Chapter 7). In the Phase 2 model, only 24 percent of the Central Lakes Deciduous South subregion is categorized as high and medium probability. Moreover, high and medium probability areas in the Phase 2 model are strongly associated with present-day water bodies. Distance to permanent lakes, rivers, and perennial streams all figure into the model. This cannot explain, however, the high and medium site potential dominating the shores of the artificial lakes south of T.H.23. This and the patterns of high and medium probability zones around other water bodies emphasize the importance of the terrain variables in the model.

Nearly all of the high probability zones in upland areas from the earlier version are absent from the Phase 2 model. A substantial part of this reduction is within areas of steep slopes. Nevertheless, there are a number of steep slopes placed within the high and medium probability areas in Phase 2. This reflects the concentration of sites on relatively higher land along the edges of river terraces and lake basins (measured by the variable height above surroundings within 90 meters). It is also significant that some of these areas also occur along older terraces and glacial outwash channels, currently some distance away from present-day water. This indicates that terrain variables may allow detection of such features even when direct data are not available.

There is also a general expansion of the high and medium probability areas in Phase 2 around the chain of mostly artificial lakes south of Richmond and the Sauk River terrace edges. This is combined with a tendency for high probability areas on the Sauk River terraces in Phase 2 to be concentrated or widest near the current river channel. However, high site probability areas have been largely replaced by medium probability around the other lakes in the region. Such changes are to be expected with the dramatic reduction in area devoted to high probability. The pattern most likely reflects terrain characteristics rather than distance to water, since the artificial lakes were not in the Mn/Model lakes database. Despite these changes, the Phase 2 model is successful in placing all sites except 21SNg into the high and medium categories. Its gain statistic is 0.70.

Return to Top

9.4 THE PHASE 3 MODEL

Figures 9.3A and 9.3B depict the Phase 3 model (T_1_3 from the Hardwood Hills ECS subsection model) of the Trunk Highway 23 corridor. Although only 19.2 percent of the entire subsection's area is classified as high or medium site potential, the area around the T.H.23 corridor has one of the highest concentrations of high and medium site probability in the model. Compared to the Phase 2 model, this latest version incorporates more of the local landscape into the high and medium areas.

In the Phase 3 model, large areas, particularly around the lakes north and south of Richmond, are classified as high and medium in site potential. The southernmost lake mapped south of Richmond (Figure 9.3A) is a permanent lake and was present at the time of the Public Land Survey. To the east of this lake, several streams feed the more extensive artificial lake to its north. This artificial lake is inseparable from an enlarged natural lake southwest of Cold Spring. It is likely that the sites proximate to the artificial lake were associated with a perennial stream or river that has since been dammed near Richmond. Because the model cannot detect the drowned water bodies, terrain and vegetation variables characteristic of these sites may be influencing the model. This particular location had a high diversity of early historic vegetation types, including an ecotone between prairie and Big Woods (Marschner, 1974). The Big Woods area, where most of the archaeological sites are found, would have both had fuel resources and have been protected from fire. This is more likely to explain the concentration of sites in this area than are the available hydrologic data.

Much more extensive areas around the lakes to the north of Richmond are assigned to the high and medium probability classes compared to the Phase 2 model. The lakes occupy a relatively flat upland area above more dissected stream valleys. This is an area of formerly aspen-oak land with relatively low vegetation diversity. Proximity to permanent lakes and terrain variables seem to both be reflected in the distribution of high and medium probability here.

Differences between the Phase 2 and 3 models are even more apparent in the area around Rockville (Figures 9.2B, 9.3B). Here, the medium and high site potential areas extend out at least one half mile from the lakes south of Rockville compared to the narrow zones of the Phase 2 model (Figure 9.2B). These lakes are in a similar situation to those north of Richmond, except that they are situated on flat land below dissected uplands and streamcourses. They are somewhat higher than the Sauk River and occupy the valley of one of the Sauk's tributaries. The high probability areas are mostly confined to this river/lake valley floor. In addition, The Phase 3 model assigns a relatively wide zone along parts of the Sauk River and its terraces to medium and high potential classes. Steep slopes assigned to zones of high and medium potential in this model mostly bound these flat valley and terrace expanses. Although procedures were implemented in Phase 3 to remove the most extensive, steepest slopes from the model probability classes, small areas and less extreme slopes may still be assigned high or medium probabilities. Again, this calls for the use of common sense in interpreting the models and applying them to practical applications.

Although the Phase 3 model predicts all of the sites predicted by the Phase 1 and 2 models, 21SNg is still not predicted. Of nine sites added to the SHPO database for Stearns County since 1995, three are not predicted by the Phase 3 model. A fourth cannot be evaluated because its reported coordinates place it outside the county. Three of the nine sites are within this local study area, and one of those is not predicted. That site is, however, in a paleovalley mapped by the DNR landforms coverage. Phase 4 models may do a better job of predicting such sites because of that additional database.

Return to Top

REFERENCES

Anderson, T.R., and M. Zelditch, Jr.
1975 A Basic Course in Statistics with Sociological Applications. 3rd edition. Holt, Rinehart and
Winston, New York.

Anfinson, S.F.
   1990 Archaeological Regions in Minnesota and the Woodland Period. In The Woodland Tradition
       in the Western Great Lakes: Papers Presented to Elden Johnson, edited by Guy E. Gibbon,
       pp. 135-166. University of Minnesota Publications in Anthropology No. 4. Department of
       Anthropology, University of Minnesota, Minneapolis.

Breakey, K.C., And C.A. Dobbs
   1995 A Phase 1 Archaeological Survey of the Proposed Northern Natural Gas Company
       Pipeline Corridor Paynesville to Watkins, Stearns County, Minnesota. Reports of
       Investigation No. 331, Institute for Minnesota.

Johnson, C.M. and C. Maki
   1996 Precontact Archaeological Investigations. In Phase I-II Archaeological and Phase II
       Architectural Investigations along Trunk Highway 23 between Richmond and I-94, Stearns
       County, Minnesota, Vol. I, edited by S. Arnott, C.M. Johnson, and J. Sluss, pp. 5.1-5.14.
       Submitted to the Minnesota Department of Transportation, S.P. 7305-83. Copies available from
       Minnesota Department of Transportation, St. Paul.

Marschner, F. J.
1974 The Original Vegetation of Minnesota. Compiled from U.S. General Land Office Survey
notes. North Central Forest Experiment Station, Forest Service, U.S. Department of Agriculture.

Murray, M.L., K.C. Breakey, and C.A. Dobbs
   1995 A Phase 1 Archaeological Survey of the Proposed Northern Natural Gas Company
       Pipeline Corridor Paynesville to Watkins, Stearns County, Minnesota. Reports of
       Investigation No. 322, Institute for Minnesota Archaeology, Minneapolis.

Return to Top

The Mn/Model Final Report (Phases 1-3) is available on CD-ROM. Copies may be requested by visiting the contact page.

MnModel Orange Bar Logo

Acknowledgements

MnModel was financed with Transportation Enhancement and State Planning and Research funds from the Federal Highway Administration and a Minnesota Department of Transportation match.

Copyright Notice

The MnModel 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.