- Management Summary
- Research Design & Time Line
- Environment & Native American Culture
- GIS Design
- Archaeological Database
- Archaeological & Environmental Variables
- Model Development & Evaluation
- Model Results & Interpretation
- Project Applications
- Model Enhancements
- Model Implementation
- Landscape Suitability Models
- Summary & Recommendations
- Archaeological Predictive Modeling: An Overview
- GIS Standards & Procedures
- Archaeology Field Survey Standards, Procedures & Rationale
- Archaeology Field Survey Results
- Geomorphology Survey Profiles, Sections, & Lists
- Building a Macrophysical Climate Model for the State of Minnesota
- Correspondence of Support for Mn/Model
- List of Figures
- List of Tables
By Elizabeth Hobbs, Curt Hudak, and Guy E. Gibbon
10 Table of Contents
10.2 Enhancing Data Confidence
10.2.1 Additional and Improved Environmental Data
10.2.2 Additional and Improved Archaeological Data
10.3 Enhancing Model Stability
10.3.1 Explicit Incorporation of Temporal Processes into the Model
10.3.2 Site Types
10.3.3 Modeling Only Surveyed Areas
10.4 Enhancing the Modeling Process
10.4.1 Advances in Software and Hardware
10.4.2 Model Regionalization
10.4.3 Modification of Modeling Procedures
10.4.4 Formal Model Classification Procedures
10.4.5 Enhancing Model Evaluation
10.4.6 Enhancing Model Interpretation
10.5 Different Approaches
10.5.1 Site Significance
10.5.2 Intuitive Models
10.6 Enhancing Functionality
10.6.1 Least Cost Path Model
10.6.2 Integrating New Kinds of Data
10.6.3 Formal Database Design
Goals for model performance were established early in the project and updated after the Phase 1 models were completed. In Phase 3, the goals were to have no more than 33% of the landscape classified as high and medium probability and to have at least 85% of the known sites correctly predicted in each region. This ratio of land area to sites predicted translates to a gain statistic of 0.61 (see Section 188.8.131.52 for a more complete discussion of project performance goals and the gain statistic).
The site probability models developed in Mn/Model Phase 3 produced gain statistics ranging from 0.44 to 0.90. They classified between 6 and 49 percent of the landscape as high and medium probability and correctly predicted 55 to 95 percent of the known archaeological sites. This summary of the results for the regional models indicates that some are not achieving the Phase 3 goals and should be improved.
Taken as a whole, however, the models performed very well with respect to the Phase 3 goals. Statewide, the gain statistic achieved was 0.76, with only 20.6 percent of the state’s landscape classified as high and medium probability to correctly predict 85.5 percent of all known sites. Moreover, the survey probability and survey implementation models provide additional information that allows distinctions to be made between low site probabilities and under-surveyed areas. Despite these strong results, there are many ways the models can be improved.
The greatest need is for improvement in data confidence and model stability, including better methods for evaluating these. In addition, the modeling process could be made more efficient, flexible, and user-friendly. This chapter discusses a variety of approaches to making these enhancements.
The models can be only as good as the data that are input. Phase 1 models were developed at perhaps the earliest possible time. Critical data layers, including National Wetlands Inventory and USGS Digital Elevation Models (DEMs), had just become available for most of the state. Other layers (historic vegetation, DNR landforms, 1:24,000 scale landforms, high-resolution soils data) were made available over the course of the project. Some of these were received too late to be included in the modeling process. Several layers were not complete statewide by the time Phase 1 ended. Because high-resolution data layers for some variables were not available, low-resolution layers were used. Table 10.1 summarizes the status of the key Mn/Model environmental data sets at the end of Phase 3.
Part of state affected
7.5 minute DEMs
Not yet available
292 quadrangles not completed in 42 counties
7.5 minute DEMs
380 quadrangles contained banding comprising 49 counties
County soil surveys
Not yet digitized
County soil surveys
Survey out of date
County soil surveys
Attribute table required interpretation in preparation for modeling
Landform sediment assemblages
All but 7 river valleys, 1 ancient lake bed/bog, and 6 upland areas
Historic cultural features (Trygg maps)
Quaternary geology (alluvium, terraces)
Soils (drainage, organic)
Historic vegetation (Marschner)
Major and minor watersheds
Minimum ½ mile distance between points
Fortunately, more high quality data have become available recently. These will allow us to discard most of the low resolution data used previously and develop better performing models in Phase 4.
10.2.1.1 Elevation Data
Because elevation data are so important to the models, replacing 1:250,000 MGC100 data with now available 7.5 minute DEMs should improve model results in the Red River Prairie and other parts of the state (see Figure 4.2 for a map of DEM status at the beginning of Phase 3). The coarse resolution elevation data that have been substituted for 7.5 minute DEMs in these areas do not represent local terrain accurately. The coarse data have the effect of smoothing the terrain, losing such details as tops and bottoms of bluffs and small-scale terrain features. Moreover, the Red River Prairie model displays artifacts of the data, in the form of high probability cells following contour lines on very flat surfaces. On the other hand, the higher resolution DEMs reflect changes in the landscape, including dikes for flood control. These may add noise to future models, confusing the analysis.
Not all of the 7.5-minute DEMs used in Phase 3 provided good quality data. Banding (see Section 184.108.40.206) is present in a number of counties, most conspicuously southeastern Minnesota. Because elevation differences between banded DEMs and USGS 7.5-minute quads are negligible, there is only a slight effect on some elevation variables. However, other variables reflect the banding and pass this trait along to the models derived from them. Fortunately the USGS, in cooperation with the state, have replaced the 7.5 minute DEMs that exhibit banding.
Because of the spotty availability of high-resolution soil data, only low-resolution data were used in the Phase 3 models. Clearly, these data are over-generalized and inaccurate for the scale of modeling being attempted. Nevertheless, these generalized data were important contributors to the models. Better soil data could provide further refinement of the patterns already detected.
Digital soil data from county soil surveys have been identified as a high priority for Minnesota. Cooperative efforts are underway at the state and federal levels to establish a program for digitizing county soil surveys (GIS/LIS News 1998). The U.S. Natural Resources Conservation Service (NRCS) currently has plans to produce high quality digital soils data (Soil Survey Geographic Database or SSURGO) data for all Minnesota counties (McCloskey 1998). These vector data are a digital version of the detailed maps found in the hard copy soil survey books published by NRCS. The dataset also includes many of the data tables found in the books.
However, simply digitizing all existing soil surveys will not provide SSURGO quality data. Twenty-one surveys, including some that have already been digitized, are not up to modern standards. A multi-year state program has been proposed to update these outdated surveys (Larson 1998). Forty-two modern surveys, as well as the 21 outdated surveys, were not mapped to orthogonalized bases, so they may not register accurately with other data layers. Research is being conducted at the University of Minnesota to develop cost-effective methodologies to convert surveys on distorted bases to geometrically correct coordinates (Bell and Krusemark 1998). Both of these efforts will contribute to the NRCS program. However, three counties have never been surveyed. It is not clear whether surveying these counties is part of the NRCS effort.
SSURGO data from NRCS are now available for 18 Minnesota counties. However, completing all 87 Minnesota counties, including those not yet surveyed, will undoubtedly take a number of years. As of this report, the contribution of high quality, high-resolution soil data to the models has not been adequately evaluated. Attempting to develop models using these data would not be efficient or effective until at least one entire ECS subsection is completed. Preferably, this subsection should have a large number of archaeological sites so that the contribution of the soil data can be fairly assessed.
In the meantime, areas of organic soils are mapped in the 1:100,0000 DNR landforms coverage. These data have a much higher resolution than the statewide soils data available previously. They will be used until even higher resolution data are available.
10.2.1.3 Vegetation Data
High-resolution vegetation data for the historic and prehistoric periods should greatly improve model performance. Unfortunately, such data will never be available. Sources of information about vegetation before the advent of aerial photographs in the 1920’s consist mostly of very low resolution maps and survey notes and pollen preserved in unique environmental settings (lakes, bogs). Barring the invention of time travel, these source data cannot be improved. However, the techniques used to collect and interpret the existing data can be improved and, perhaps, be used to develop GIS layers that can contribute better information to the models.
The best sources for historic data in Minnesota are the Public Land Survey System (PLSS) plat maps and surveyors notes. These are the primary sources for the Marschner map of historic vegetation, the DNR bearing trees database, and the Trygg maps. All of these datasets could be improved by judicious use of the source material.
Marschner (1974) and Trygg (1965-1969) generalized the data mapped by the surveyors to 1:500,000 and 1:250,000 scales respectively. Moreover, their map bases are almost certainly not georeferenced to accurate coordinates. The PLSS plat map features could be georeferenced to accurate section corners and digitized at their original scale (about 1:31,680), showing features such as lakes, rivers, streams, wetlands, roads, trails, dwellings, and other cultural sites. Although much of the information in the interior of sections would be interpolation or conjecture, the resulting maps would include a higher level of detail and more features than the Marschner or Trygg generalizations.
The PLSS surveyors’ notes also contain information about vegetation composition and structure that can be extracted and used to attribute, and sometimes add features to, the digitized plat maps. This includes the locations and descriptions of bearing trees that were used to mark section and half-section corners, as well as junctures with lakes and streams. Surveyors’ notes also include descriptions of vegetation composition and structure along survey lines where significant changes occur ("leave forest, enter prairie"). This information is incompletely captured in the DNR bearing trees and survey point vegetation GIS layers. The general descriptions from the survey corners have been captured in the DNR database, but these are attached to the section corners, not extrapolated along the survey lines they are intended to describe.
The combination of surveyors’ notes and plat maps can be used to develop GIS vegetation and cultural features layers at a higher level of resolution than those now available. The procedure is very time-consuming, but it has been completed for a number of Minnesota State parks (Boudreau and Hobbs 1994). Expanding this effort to the entire state would be expensive, but could greatly enhance our understanding of the presettlement landscape.
Trygg (1964-69) maps capture some of these features, but at a very coarse scale. William Trygg copied features from the surveyors plat maps to base maps at a scale of 1:250,000. It is not clear how much use he made of the survey notes. The Trygg maps show boundaries between vegetation types, such as prairie and woodland, but do not distinguish between different types of woodland. According to LMIC, Trygg maps have been digitized for the state, but are not yet available for distribution. Where data from these maps were included in Phase 1 models, the models improved. The most significant variables were distance to grassland and distance to woodlands. However, vegetation diversity derived from these maps also contributed to models. When these maps become available, their use could enhance the models at less cost than more accurate mapping from the PLSS source data. By combining the more accurate vegetation boundaries from Trygg with the better defined vegetation types from Marschner and the public land survey corner data, it may be possible to improve the value of the information obtained from the three sources separately.
Vegetation maps for prehistoric periods would almost certainly improve the ability to model sites before the early contact period. However, it is unlikely that such will be available at a scale that will contribute to the models. Most paleovegetation mapping is very coarse scale because the pollen record on which it is based is sparse (see Section 6.3.3). Pollen is best preserved in situations where organic decomposition is absent or very slow, such as shallow lakes, swamps, and bogs. Such sites are not uniformly distributed across the state. The pollen record is more often used to infer sequential vegetation change in the area surrounding the pollen source, and even this application has its limitations (Kellman 1980, p. 129-136). Moreover, the pollen assemblage may not be representative of the local landscape. Because depositional sites are often in wetlands, the assemblage may be dominated by swamp species that are not typical of the surrounding area. Since pollen may be carried very long distances by wind, pollen from tall wind-pollinated species is widely dispersed and may be over-represented in pollen assemblages. Differences between species in rates of pollen production and preservation also affect assemblages, with the result that important local species may be missing from the record.
Because of the inherently poor quality of data for historic and prehistoric vegetation mapping, consideration should be given to modeling the desired information. Modern elevation, soils, and hydrologic data, as well as additional historical documentation, could be used in conjunction with digitized PLSS plat maps to adjust lake and wetland boundaries and stream courses and to add features missing from the interiors of sections. This would produce a higher resolution model of the historic landscape that may prove useful as a component of an archaeological predictive model. Although some components of this modeling process would necessarily be intuitive, given the gaps in available data and in our understanding of historic vegetation landscape relationships and dynamics, much of it could be codified into decision rules that would be completely replicable from one place to another.
It may be possible to use a combination of paleoclimate data, pollen records, geomorphology, terrain, hydrology, and historic vegetation boundaries to model vegetation change backward through time. However, this modeling process would be highly intuitive. The resulting models would necessarily be at a coarser scale and would become less reliable as they reach further back in time. However, they may be more capable than modern data for supporting archaeological predictive models of paleoindian and archaic sites.
Higher resolution geomorphologic data should improve the predictive model. For this project, landform sediment assemblages (LfSA) were mapped at a scale of 1:24,000 for seven river valleys, one ancient lake/bog, and six upland areas comprising approximately 13 quadrangles in the state (Figure 6.2 and Chapter 12). This layer also provides an assessment of the suitability of each LfSA unit for containing archaeological resources. These assessments are difficult to test, however, since most previous archaeological surveys have not adequately evaluated the subsurface for cultural resources. Unfortunately, the LfSA layer covers only a very small portion of the state’s area, though mapping continues. The map scale is undoubtedly the most appropriate for the high resolution models being developed for this project. Variables from preliminary landform sediment assemblage maps (without suitability ratings) were incorporated into a Phase 1 model, with the result that the variable landform diversity improved the model. However, it will be impossible to adequately assess the utility of these data for modeling until a large contiguous area is mapped (preferable an entire ECS subsection). It may be possible to do a preliminary test in one or more of the largest river valleys mapped, provided enough sites are recorded within the mapped area to support the statistical analysis.
There is an approach which would allow the incorporation of the landform sediment assemblage layer into the models, even though only a small area will be affected. The landscape suitability rankings that were interpreted from the LfSA data can be incorporated into the existing statistical models by overlay. This would produce a composite model, along the lines of the survey implementation model, that provides new information for those areas that have been mapped by the project geomorphologists.
In addition, the 1:100,000 scale digital landform maps from the Mn/DNR are ready to be used for modeling. At the end of Mn/Model Phase 3, the original attributes for this layer were interpreted to establish correspondence between the data as mapped and the Mn/Model classification system developed for the landform sediment assemblages (Section 220.127.116.11). New attributes consistent with the landform sediment assemblage classification were joined to the DNR landform coverage feature attribute table. These attributes will support modeling, as well as allow integration with the higher resolution LfSA data in the future. Models using variables derived from this data layer should be a high priority, as the variables on terraces and on alluvium, both derived from a MGC100 1:100,000 source, figured into many models. Mining areas that were excluded from modeling were derived from the same source and could be refined from this new layer.
The attribute table that accompanies this layer (combining the original DNR attributes and additional attributes from the Mn/Model landform sediment assemblage classification system) contains eight fields containing potentially useful information, in contrast to the single attribute of the MGC100 Quaternary geology layer (see Appendix B). This will support the derivation of additional variables. Moreover, more feature types, such as beach deposits and glaciofluvial valleys, are mapped. Although Mn/Model detected some of these features by interpretation of other data sets, such as elevation, these data should improve the ability to model and interpret high probability areas that may be associated with these features. It may also reduce the extent of other rugged terrain being misinterpreted as suitable for archaeological sites. These data have the potential to significantly improve the predictive models for surface sites until a high resolution geomorphic context can be developed for Mn/Model.
10.2.1.4.1 Developing a Geomorphological Context for Mn/Model
At a higher level of abstraction, the geomorphological data that are now available and will become available in the future can be used to develop a geomorphological context for Mn/Model. Geomorphic processes of landscape evolution can affect the surface visibility, integrity, and survival of archaeological sites. While some sites may be obscured by sediment deposition, others may be destroyed or deflated by sediment erosion (Hajic 1990a-b; Holiday 1992; Lasca and Donahue 1990; McManamon 1984; Needham and Macklin 1992; Stafford 1995; Stein 1987; Stein and Farrand 1985; Van Nest 1993). Failure to consider these geomorphic processes could lead to the development of distorted or incomplete predictive models (Warren and Asch 1996). Mn/Model's predictive models will be incomplete until these processes are taken into account.
The ability to locate sites of a particular cultural period depends on our ability to locate the landforms and depositional environments where sediments of the appropriate age are preserved. Until these environments are mapped, Mn/Model’s assignments of high, medium, and low probability classes will be most accurate for relatively recent, i.e. Woodland, sites. At the very least, future models will incorporate information about environments where surface sites could not possibly have been preserved. At best, the ability to predict locations of buried sites should be enhanced. This will require a better understanding of the geomorphological context of buried sites in Minnesota.
Most archaeological surveys in Minnesota have not systematically recorded a range of significant geomorphic observations necessary to develop geologic contexts where buried sites are most likely preserved. Phase I archaeological reconnaissance is normally conducted using conventional surface survey methods, such as visual inspection of exposed ground surfaces and digging shallow shovel tests in areas covered with vegetation. Little effort is made to search for buried cultural deposits, except to examine exposed stream cutbanks, tree throws, and rodent burrow spoil piles. Even in Phase II archaeological evaluations, excavations are rarely deeper than 50-100 cm, and hand mechanical probing seldom reaches deeper than 1.5-2.5 m. However, in Minnesota, as in adjacent areas such as the northern Plains (Artz 1996:387), much of the archaeological record, and in particular its earlier part, is buried and undetectable using conventional methods of pedestrian reconnaissance and shallow probes.
A better-developed geomorphological context will improve the predictive abilities of the models and enhance the implementation of Mn/Model in cultural resource investigations. The development of this context is a three-step process:
1. Identify, define, and delineate in a GIS layer the major landform units (geomorphic features) of the physical landscape of Minnesota.
2. Reconstruct the geomorphic development of these landform units, including their materials, material sequences, age, depositional, and post-depositional environments.
3. Determine the archaeological significance of these geomorphic features, with special attention paid to the identification of buried horizons and to the location of previously unknown archaeological sites. Assign landscape suitability rankings based on these determinations. This is, in essence, a geomorphological model of landscape suitability for preservation of both surface and buried sites.
This procedure is more completely described in Chapter 12. A previously published example of such mapping, albeit a different style, can be found in Bettis and Hajic (1995). Such models are being developed from the landform sediment assemblage mapping accomplished as part of this project. Figure 10.1 illustrates the results. However, these maps cover only a very small portion of the state (Figure 6.23).
Unfortunately, the DNR landform coverage lacks the information content and spatial detail required for such model development. The Mn/Model geomorphologists have interpreted this map to provide MnDOT and SHPO with a GIS layer showing mostly thicker Holocene-aged sediments. However, this coverage does not really address either depositional environment nor landscape suitability because of the coarse scale.
The extension of the landform sediment assemblage mapping to a statewide GIS layer is likely to take years. Its development may be prioritized to first focus on places where Minnesota’s landscape was severely altered or where massive development is projected. For this reason, seven river valleys and one ancient lake bed were selected for mapping as part of this project. Another option would be to perform such mapping in the early planning phases of MnDOT projects. The resulting landscape suitability model can be interpreted in conjunction with the archaeological predictive model for the project area only, thus informing the design of the project survey strategy. This latter method and implementation strategy has currently been adopted by MnDOT.
The remainder of this section is devoted to descriptions of two specific situations (Glacial Lake Duluth and two glaciofluvial valleys), followed by descriptions of three periods of dramatic landscape modification. These are provided to emphasize the magnitude of geomorphic changes in Minnesota and their effect on the visibility and preservation of archaeological sites.
Glacial Lake Duluth
A theoretical basis in the predictive modeling of hunter-gatherer settlements is that they are located near water. One would expect, then, that sites of all periods be clustered around the Lake Superior Basin. However, sites along the southern shoreline dating before ca. 3000 B.C. were destroyed or deeply buried by a variety of geological processes, including:
o The Marquette re-advance (a Laurentide lobate surge) at ca. 7900 B.C., with the formation of Glacial Lake Duluth
o Catastrophic discharges from Glacial Lake Agassiz and large-scale storms that produced short-lived catastrophic transgressions along the shoreline between the Minong and Houghton Lake levels (7500-6000 B.C.)
o A rise in water level, known as the pre-Nipissing transgression, at ca. 3000 B.C. (Clayton 1983; Drexler, Farrand, and Hughes 1983; Phillips 1993; Teller and Thorleifson 1983).
Furthermore, differential isostatic deformation moved sites of the same age along the shoreline to increasingly higher altitudes northwards along the shore. Consequently, elevation alone cannot be used as a guide to the location of sites. Moreover, these catastrophic transgressions most likely changed the location of other attractive settlement localities, such as marshes and bogs. Post-Nipissing lake-level fluctuations, bio- and cryoturbation, the aggregate industry, and urban and town development are more recent processes that have contributed to the burial or destruction of archaeological sites along the Lake Superior shoreline.
Phillips (1993) has outlined a space-time model for the distribution of shoreline archaeological sites in the Lake Superior Basin that could be integrated into a GIS layer of landform-sediment assemblages for this critical area of the state.
Glacial Rivers Warren and St. Croix
Periodic, vigorous surges of glacial melt water through both the Glacial River Warren and St. Croix valleys most likely buried and destroyed archaeological sites in the river valleys before ca. 7500 B.C. and possibly one millenium earlier (see Chapter 12). At this time they ceased to serve as the outlets for Glacial Lakes Agassiz and Duluth, respectively (Drexler et. al. 1983; Matsch 1983). With the termination of the deep erosion caused by the glacial rivers, natural dams formed across both the Mississippi and St. Croix rivers, which occupy the Glacial River Warren and Glacial St. Croix valleys, as the rate of their flow decreased and the rate of sedimentation rapidly increased at the confluence of trunk and tributaries (Eyster-Smith et al. 1991; Wright et al. 1995). The result in the St. Croix River was the formation of Lake St. Croix, which formed when a natural dam (an alluvial fan) was created across its mouth by the Mississippi River near Hastings. Deposition of an alluvial fan across the River Warren trench near Wabasha by the Chippewa River created Lake Pepin. Since their formation at ca. 7500 B.C., both lakes have prograded downstream. For example, Lake Pepin, which is now ca. 35 km long, was initially 120 km long and stretched from the Chippewa River fan upstream to St. Paul.
The formation and transformation of these lakes has undoubtedly affected the location and visibility of archaeological sites along these sections of the rivers. Two examples illustrate the magnitude of these changes. At the height of the Mississippian occupation of Minnesota, major Mississippian village sites were seemingly clustered at the juncture of the Mississippi and Cannon rivers, with the juncture a determining variable. However, a reconstruction of the history of Lake Pepin demonstrates that these villages were actually sited at the juncture of the Cannon River and Lake Pepin. The northern end of Lake Pepin is now a dozen miles or so to the south. As a second example, as the lakes formed they gradually raised the water table of the immediate floodplain. Coupled with sporadically severe upland erosion, sites on alluvial fans, such as King Coulee (Perkl 1996), became deeply buried, with their lowest and earlier horizons below the floodplain. Geomorphic maps of Lake Pepin, outlining the distribution of relevant landform-sediment assemblages, are being developed by the U.S. Army Corps of Engineers. The Lake St. Croix maps were produced as part of the Mn/Model geomorphologic investigation (Chapter 12). Similar types of mapping should be completed in other areas of Minnesota.
The Mid-Holocene Dry Period
One of the most disruptive climatic episodes in the past as far as the archaeological record is concerned was the mid-Holocene dry period. Because of the magnitude of this disruption, the mid-Holocene has been a focus of geoarchaeological study in North America (Artz 1995; Bettis 1995; Bettis and Hajic 1995; Meltzer 1995; Root and Ahler 1987).
In Minnesota, most large-scale changes in settlement strategy before ca. 1000 B.P. Were probably related to significant environmental changes. As a framework for discussion, the varve sequence from Elk Lake (Bradbury and Dean 1993a) in Itasca State Park can be used to establish dates for subdivisions of the Holocene, although transitions between subdivisions tend to occur at slightly different times in different areas of the state. The following divisions and their characteristics have been derived from the varve sequence (Bradbury and Dean 1993b):
Late Glacial: 11,600-11,000 B.P. Conditions much colder and drier than at present; spruce (Picea) dominates the pollen record.
Late Glacial/Early Holocene Transition: 11,000-10,000 B.P. Marked increase in temperature and precipitation; global warming at the end of the Pleistocene.
Early Holocene: 10,000-8500 B.P. Relatively cool and moist conditions when pine (Pinus) dominates the pollen record. The period was warmer and moister than preceding times, and a little cooler than but about as moist as at present.
Early-Middle Holocene Transition: 8500-7800 B.P. Change from forest to prairie.
Middle Holocene: 7800-4500 B.P. Low annual precipitation and a general increase in July temperature. Drought conditions and strong winds were widespread in Minnesota between ca. 8000-4000 B.P.
Middle-Late Holocene Transition: 4500-3500 B.P. Forest increases. After 4000 B.P., the climate at Elk Lake was dominated by a tropical air stream during the summer and dry arctic and pacific air streams during the winter. Large-scale variations in climate cease, although decadal and multi-decadal variation occurred, such as the medieval Warm period and the Little Ice Age.
Late Holocene: 3500 B.P. to present. Characterized by modern vegetation and climate.
As the climate became drier in the Early-Middle Holocene transition, vegetation responded rapidly and prairie spread up into the northeastern corner of Minnesota (Bartlein and Whitlock 1993; Dean et al. 1996; Forester et al. 1987; McKinnon and Stuart 1987). Besides major vegetation shifts, sedimentation increased rapidly and westerly zonal winds dramatically increased dune activity in areas such as the Anoka Sand Plain (Fritz et al. 1991; Grigal et al. 1976; Keen and Shane 1990). At Elk Lake, by the end of the Mid-Holocene prairie period, annual precipitation was 100 mm lower and July temperature about 2 degrees C warmer than today (Bartlein and Whitlock 1993), although it was not necessarily warmer in all areas (Dean et al. 1996). Numerous studies in Minnesota trace the process of aforestation at the end of the mid-Holocene prairie period (e.g., Almendinger 1992).
Among the changes that most affected settlement location was a general shift from forest to prairie, drastically lower lake levels (Almendinger 1993), and increased sedimentation (Almquist-Jacobson et al. 1992). Archaeologists have argued whether the northeastern Plains were only sparsely occupied by Native Americans at this time or if sites are masked from view by sedimentation (Artz 1996; Reeves 1973; Sheehan 1995). Increasing numbers of excavations (e.g., Running 1995) indicate that many sites are buried in alluvial fans and to a lesser extent in colluvial deposits in upland swales. Considerable effort is being expended at the present time by geoarchaeologists in modeling the distribution of landform-sediment assemblages that might contain archaeological components of this time period (Bettis 1995).
The Little Ice Age
Minnesota’s landscape changed severely during periods of human occupation both during and prior to the mid-Holocene. Late Holocene changes were also severe enough to cause shifts in settlement locations, or to destroy or mask sites. Examples are the Little Ice Age and Euro-American settlement. The Little Ice Age is discussed in this subsection and the impact on archaeological sites of Euro-American settlement in the following subsection.
Climate in Minnesota and the broader north-central region of the United States is determined by the interplay of three air masses that join over the state. These are dry Pacific air from the west, a cold, dry arctic air stream from the north, and a warmer, wetter tropical Atlantic air stream from the south (Wright 1976). Changes in the relative dominance of one air mass or another are thought to be responsible for shifts in Holocene climate and therefore vegetation in the state. For example, the eastward expansion of prairie during the Middle Holocene is thought to be a result of an increase in the westerly flow (Webb et al. 1983). The Little Ice Age was a cool, moist interval between ca. A.D. 1500-1850 (500-150 B.P.) in many parts of the Northern Hemisphere (Bradley and Jones 1992; Fritz et al. 1994; Grove 1988). It was this cool, moist interval that apparently lead to the replacement of fire-susceptible hardwoods (oak savanna) by mesic hardwood forest (Big Woods) in southeastern Minnesota (Grimm 1981) and more frequent large floods in southwestern Wisconsin (Knox 1993). Flooding and the expansion of Big Woods vegetation most likely affected site location.
The complexity of mapping the environmental change caused by the Little Ice Age is demonstrated by data from lakes in northwestern Minnesota and Northeastern North Dakota. At Elk Lake in Itasca State Park in northwestern Minnesota, there is a good correlation between lake-level and moisture availability in the lake’s uninterrupted sequence of varves that extend back into the Early Holocene (Bradbury and Dean 1993a). Although part of the period is moister and cooler, there are dry intervals from ca. A.D. 1690-1750 (310-250 B.P.) And A.D. 1830-1860 (170-140 B.P.). At Devils Lake in northeastern North Dakota, there is evidence of an arid climate and drought conditions during the Little Ice Age that were as extreme as during the Dust Bowl of the twentieth century (Fritz et al. 1994). In the Great Plains, such periods of aridity result from high-pressure systems that divert the flow of summer moisture from the Gulf of Mexico away from the Plains.
This example illustrates the existence of spatially variable climatic patterns in the Minnesota area during the Holocene that resulted from the interplay of the three dominant air masses over the state, a conclusion also reached by Baker and his colleagues (1992). The Little Ice Age was not uniformly moist and cool, but apparently exhibited very steep climatic gradients between the Great Plains and regions to the east. While site density may have increased in the emerging Big Woods area of southeastern Minnesota, sites farther west may be buried in sediment in former lake beds because of a greater erosion rate. Sites farther west may also be more rare in general than in earlier and more recent periods in the Late Holocene. Unfortunately, the Mn/Model paleoclimate model may be too coarse to map these precipitation clines from east to west across the state. The SHPO archaeological database can be examined for disparate relative density of sites of this age in the southwest compared to the southeast section of the state, if enough sites from this period can be identified with confidence. Evidence to interpret this regional disparity should also be recorded in the core profiles provided by the Mn/Model geomorphology team. If site locations are being masked by geomorphic processes in the southwest section, then this gap in the archaeological database could be filled by a geomorphic model that identifies potential for buried sites.
During the historic period, ca. A.D. 1880-1930 (120-70 B.P.), intense disruption of the natural landscape and the archaeological record occurred in Minnesota. The main destructive processes were logging, mining, and dam construction in the north and forest clearance, intensive cultivation, channel widening, and drainage of wetlands in the south. The processes severely affected hydrology, slope processes, vegetation, sedimentation, and water quality, as well as the visibility and survival of archaeological sites. For instance, the majority of sites around some large dammed lakes, such as Big Sandy Lake in Aitkin County, have been destroyed. Of all landscape areas, historic floodplains were most severely altered during this period (Baker et al. 1993; Knox 1987). Extensive soil erosion resulted in increased runoff, more frequent flooding, and the replacement of native lowland and upland plant communities with ruderal (disturbed ground) vegetation. Accumulations of sediment derived from uplands and slopes reached 2 m thick in channel belts and 1 m thick in the floodplains. Not only were lowland sites buried, many upland sites were deflated or completely destroyed.
Archaeologists have not yet assessed the impact on the archaeological record of these processes of historic disruption. This period of historic disruption affects Mn/Model in several ways:
(1) Modern lake levels and river valley configurations have changed.
(2) The distribution and composition of vegetation communities has been severely altered.
(3) Wetlands have been drained.
(4) Sites have been buried, eliminated, or had their context destroyed
(5) The universe of archaeological sites in many parts of the state has been truncated, so that significant segments of settlement models may be absent.
Many other natural processes that impacted subsistence and settlement patterns in precontact Minnesota are part of a landscape evolution context for Mn/Model, such as Late Holocene valley incision (Mooers and Dobbs 1993) and significant dune movement within the last 1000 years (Dean et al. 1996:151). Examples of geomorphological contexts for cultural resource management already exist for some regions of bordering states, including Iowa, Illinois, and Kansas (Benn 1990; Bettis 1990; Bettis and Benn 1984; Hajic 1990a-b, 1993; Mandel 1992, 1995; Stafford et al. 1992; Van Nest 1993).
The Minnesota SHPO archaeological database has proven to be the most significant weak link in the modeling process. Minnesota’s low precontact settlement density and dynamic postglacial environment both contribute to the problem of the sparseness of sites. The very low numbers of sites from probabilistic surveys in many regions dictated that Phase 3 models be developed using all known sites. Even so, nine of the 27 ECS subsections had fewer than 100 sites available for modeling. Without a larger probabilistic database for comparison, it is not possible to determine the impact this has on the models.
In general, larger databases produce better, more reliable models. Poor model performance may, however, be masked by the effects of survey bias. For example, the Littlefork Vermilion Uplands subsection contains only 25 known sites. However, the model built for that subsection predicts 92% of these sites in 18.32% of the landscape, for a gain statistic of 0.80 (see Section 8.18.2 for a complete discussion of this model). On the face of it, this is stellar performance. However, the low Kappa statistic value for the preliminary models for this subsection warns that the model is not stable. Moreover, the number of negative survey points in the subsection is low (207) and the survey probability model performs nearly as well as the site probability model, implying a strong degree of survey bias. These factors must be taken into consideration in the implementation of the model.
Counties and regions with poorly developed archaeological data can be readily identified using the archaeological database and survey probability models. Although two regions may have a similar number of archaeological sites, one may have a greater need for more surveys because it is poorly known, as evidenced by the number or distribution of previous surveys. Other regions may have very low numbers of both known sites and surveys. In contrast, regions with relatively few sites but a large number of surveys, which have been distributed widely throughout the landscape, may simply have low site densities.
The most effective way to improve models for subsections with low site numbers is to add more sites to the archaeological database. This should be done, however, using probabilistic sampling designs that will reduce the level of surveyor bias in the database. In addition, the archaeological database suffers from inconsistent quality attributable to survey techniques, the accuracy of the data recorded, and the way in which sites have been represented in the models. These issues, and suggestions for remedying them, are addressed in the following subsections.
10.2.2.1 Improved Sampling Design
Most data in the archaeological database came from non-probabilistic surveys. These data have been collected over many years by a large number of professional and amateur archaeologists. These surveyors have tended to concentrate their efforts in places they expected would contain archaeological sites.
Because of the low site numbers problem, Phase 3 models were built using the entire archaeological database (with the exception of isolated find sites represented by single artifacts only). This incorporates into the models the biases inherent in the non-probabilistic and stratified probabilistic surveys (see discussion below) that produced the majority of the data. Because of these biases, many environments in the state have been poorly surveyed (Figure 8.6).
This situation was brought to light in Phase 1 by an examination of the negative survey data. The Mn/Model archaeological database includes nonsites, as represented by random points taken from surveyed areas where no sites were found. The surveyed places mapped were from stratified probabilistic or qualified CRM surveys. These were determined not to be representative of the environment as a whole, as is apparent from examining the distribution of negative survey points between the high, medium, and low probability areas in the model evaluation tables in Chapter 8. If these points were statistically random, their distribution should reflect the distribution of cells within their region. In most cases they do not. More often, negative survey points are distributed more like known sites than like the total population of cells.
This finding supports the assumption of survey bias in the stratified and CRM surveys, in favor of places where archaeologists expect sites to occur. Consequently, we must assume that models based on this database are biased in a similar way. The survey probability and survey implementation models (Chapter 8) developed in Phase 3 illustrate the geographic pattern of this bias.
Future enhancements to Mn/Model will rely on the foundation of a more complete database, with the goal of categorizing less of the landscape as "unknown." This implies more sites, more negative survey points, and a more representative sample of the landscape. To strengthen the archaeological database for statistical modeling, future surveys should be subject to rigorous, probabilistic sampling design. This is the only way to be assured that the sample being modeled is representative of the population of sites as a whole.
10.2.2.1.1 Probabilistic Sampling
For any kind of statistical modeling, it is best to have a very large, consistent, representative database. This database should be derived from probabilistic, preferably simple random, samples of the landscape, since simple random samples of archaeological sites are not possible (Kvamme 1990, p. 285).
The Minnesota archaeological database is not the result of one coordinated, probabilistic survey. In fact, only a very small portion of the database was collected using anything approaching probabilistic methods. Only three counties were sampled using truly random sampling procedures (see Appendix C). All other surveys considered to be probabilistic were sampled using some kind of stratification procedure (near water, away from water) or as part of a trunk highway or pipeline survey.
Stratified surveys have the potential to bias the database, even if sampling within the strata is truly random. Stratification makes certain that all landscape elements (as defined by the researcher) are sampled. For a truly probabilistic survey, these should be sampled in proportion to their representation in the landscape. This rule of thumb has not always been followed in previous stratified surveys. For the Minnesota Statewide Archaeological Survey and the 1995 Mn/Model survey that was based on it, the primary stratification was "near water" and "away from water." There were many landscape elements identified and sampled in the "near water" category. Uplands away from water, which are more extensive, were less heavily sampled. Archaeological sites tended not to be found in the uplands. This may be because they are not there, or it may be because the uplands were under sampled.
For 58 counties, the only "probabilistic" data came from trunk highway and pipeline surveys. Trunk highway and pipeline surveys, where the entire corridor was surveyed, do not contain the same kind of bias as the stratified surveys described above. However, they do pose two other potential problems. First, the highway or pipeline corridor may have been selected because of certain landscape characteristics that may not be representative of the region as a whole. Second, with a large number of adjacent 40 acre parcels sampled in a corridor, problems of spatial autocorrelation are more likely (Kvamme 1990 p. 285).
Any future surveys have the potential to add new sites to the archaeological database. Only probabilistically based surveys have the potential to add sites and negative survey points that are truly representative of the diverse environments of Minnesota. The implementation plan developed by MnDOT for this project (Chapter 11) suggests the adoption of a stratified random sampling technique tied to the survey implementation models. To be truly probabilistic, the number of points surveyed in the high, medium, and low/unknown parts of the project area should be proportional to the areas of these model classes. However, fiscal considerations and the need to be certain that significant cultural resources within the project area are not overlooked will dictate that high probability areas are sampled most intensively. This strategy risks perpetuating exactly the same kinds of survey bias that already exist in the archaeological database and that have been captured in the models. It will be up to MnDOT’s Cultural Resource Unit to make certain that enough sampling is done in the medium and low/unknown probability categories to begin to offset this bias. Quantitatively, this will be an uphill battle.
10.2.2.1.2 Improved Field Methods
The likelihood of an archaeological site being discovered in a survey depends on the field methods used, as well as the experience and abilities of the field crew. Generally, older surveys are considered less reliable than more modern ones because they were based on less scientific field procedures. To further complicate matters, even within counties that were part of a modern MnSAS survey, there were differences in field methods that may have affected the probability of finding sites if they were present (see Section 5.3). Survey quality probably varied between different surveying teams. Consequently, some places that were recorded as negative survey points may in fact contain archaeological sites that were not discovered. Some standardization of field techniques for future surveys, as part of the Mn/Model implementation plan, could improve the likelihood of finding sites and enhance the representativeness of the archaeological database. These may include:
- Procedures for contacting landowners; since denied access (or failure to make contact) may make it impossible to survey assigned random points.
- Rules governing when an assigned parcel may be deemed unsurveyable.
- Unambiguous replacement procedures for selecting a new sample location when an assigned parcel cannot be surveyed.
- Maximum allowable spacing of transects and shovel tests.
- A threshold minimum acceptable surface visibility for pedestrian survey.
- Rules governing artifact collecting between transects.
- Depths and diameters of shovel tests.
The procedures adopted for the 1996 Mn/Model field surveys (Appendix C) could serve as a starting point. Standardized, high quality procedures accepted and adopted by the archaeological community could improve both the probability of finding sites where they are present and confidence in the negative survey data, where sites are reported to be absent.
10.2.2.1.3 Systematic Recording of Valid Nonsite Areas
An important component in building statistical predictive models is the use of nonsite control locations for comparison with site locations. Mn/Model has had to choose between two problematic alternatives. The first was selecting points within surveyed sections that reflect biased survey procedures – and as a result do not truly represent the environmental diversity of our study areas. The second employed truly random points that have not been surveyed (except fortuitously) for the presence of sites.
Since Mn/Model did not have access to digital data for locations of all past surveys, the nonsite database used is incomplete. SHPO has plans to digitize all surveyed areas for which they have reliable maps. Completion of this database should help improve future models, although the same survey biases are likely to be apparent. At the same time, the reporting of survey and excavation results should be improved. This problem is addressed for future surveys by the Mn/Model implementation plan (Chapter 11).
10.2.2.1.4 Filling Gaps in the Archaeological Database
During the 12,000-year history of the state's contact and precontact Native American societies, populations were always sparse compared to Midwestern states to the south. These populations were organized, too, in a variety of subsistence-settlement patterns that all included some degree of seasonal mobility. Preferred site locations were diverse, and their locations are masked in many cases by recent and dramatic changes in the landscape. In fact, the archaeological record is a product of numerous cultural and natural site formation processes (Schiffer 1987). Consequently, parts of the archaeological record are under represented or not represented in the Mn/Model archaeological database. These are likely to be older sites, which have been buried or destroyed by geomorphic and anthropogenic processes. Since a statistical model will best predict the most abundant types of sites in the database, Mn/Model is best for predicting the location of more recent sites, primarily Woodland to Historic.
There are different approaches to determining which portion of the archaeological record is under represented or not represented in the Mn/Model archaeological database. An example is Hammer's (1993) inductive statistical approach. Site types within broader adaptive types are identified and compared to ethnographic analogues to determine which kinds of sites are most likely missing (and presumed buried, not recognized, or destroyed). Archaeological surveys focused on locating such sites may produce data that can be incorporated into future predictive models. However, this would be at the expense of biasing models towards those types of sites that were the focus of the surveys.
As an alternative, systematic scientific surveys targeting the kinds of places where sites, regardless of their type, could be buried can be conducted. This shifts the focus from locating sites to exploring landscape units. Still, it promulgates the same kind of survey bias that plagued most previous surveys in the state. However, concentrated survey efforts for buried sites could improve the the database by adding a depth dimension and provide a test of the geomorphic models.
Finally, the only method available for adding archaeological data in an unbiased way is a simple random survey in each county. However, given the sparse site dilemma, a very large number of random points would have to be surveyed to find even a few significant archaeological sites. This would be a very expensive undertaking. A possible compromise would be to focus new surveys in areas currently modeled as having low potential for having been surveyed. Even if particular environments or landforms are targeted for survey within these areas, the net result will be reducing the overall bias of the database.
10.2.2.2 Accurate and Complete Site Records
Aside from the quality of field survey procedures used, how well archaeologists recorded their survey results can greatly affect the quality of the database. Issues include the completeness and reliability of the site record, defining a site, determining and recording site locations and boundaries, and use of consistent terminology. Of these, inaccurate recording of site location has probably had the greatest effect on the current models. Completeness and accuracy of the other aspects of the records may determine what future modeling enhancements are possible.
Accurate site locations are necessary for evaluating associated environmental variables. If sites are not accurately located, then false relationships between sites and their environment are built into the models. The accuracy with which site locations were recorded in the Mn/Model archaeological database varied considerably. At the beginning of the project, sites to which the SHPO had assigned low locational confidence were dropped from the database (Section 5.6.1). However, an analysis performed in Phase 3 discovered even more previously undetected problems in recorded site locations (Section 5.2 and Chapter 12).
Two strategies will be needed to rectify this situation. First, quality control should be performed on the existing database. Inaccurate site coordinates should be corrected. If this is not possible, the site in question should be coded as "low locational confidence." Second, measures to improve the recording of site coordinates in future surveys should be adopted. These should include standardized, high quality procedures in which field crews are trained. Components of these procedures should include:
- Rules for defining and mapping site boundaries.
- Rules for determining a point to record as a site centroid (to insure, for instance, that it is not in water or otherwise outside of the site itself).
- Prohibition from using photocopies of measuring tools.
- Encouraging the use of Global Positioning System (GPS) technology and orthophoto quads to improve mapping accuracy.
Other aspects of the archaeological database that are important to future model enhancements are cultural context, chronology, and function. For many sites, these are not recorded because they can’t be determined from the artifacts or site context, because the artifacts were not adequately analyzed, or for various other reasons. In other cases, they may have been inaccurately interpreted and recorded. Improving the quality of these components of the archaeological record will provide more records that can be used to stratify data into temporal, functional, and cultural categories. If sufficient numbers of sites are present in the resulting groups, it should be possible to build more precise models (see Section 10.3.1).
10.2.2.3 Alternative Representation of Archaeological Data
Cultural resources are distributed in varying densities across the landscape. Site centroids are not truly representative of this distribution. Modeling sites as centroids erroneously classifies many cells where archaeological resources are present as having no archaeological resources. Representing archaeological resources as site polygons or as continuous distributions of artifact densities are two alternatives to reduce this error.
Representation of sites as polygons requires the identification of site "boundaries" within the continuous distribution of artifacts. One solution is to decide the degree to which a resource must be present, then draw a line around polygons that contain the required amount of the resource. This method is used, for example, in vegetation mapping, where forest may be defined as having 75 percent tree cover, with lines drawn around all areas with 75 percent tree cover or greater. Archaeological sites have traditionally been mapped by this means, though decision rules about what density of artifacts is required to define a site have not been formal or consistent. Because of the low confidence in the delineation of site boundaries in the Minnesota archaeological database, the SHPO Archaeologist advised the Mn/Model Team against including them in the GIS for this project. However, SHPO plans to digitize known site boundaries in the future. When these polygons are available, they will be used for modeling. They will represent, at the very least, the minimum limits of many sites.
Site densities, the number of sites in a unit land area, are an alternative way to represent site distribution across the landscape. However, this approach requires the use of a larger cell size (several square kilometers) to obtain sufficient variability in the dependent variable (Kvamme 1990, p. 270). This model resolution would not be adequate for the planning applications for which Mn/Model is intended. However, such a surface of site density, or a surface of the ratio of sites to negative survey points per unit area, could be used in conjunction with a high resolution model for relative weighting of model values between regions.
The ideal approach to representing archaeological resources would be as a grid of cells with values for the artifact density assigned to each cell. This could be accomplished by assigning artifact density values to cells for which data are available, then interpolating the values of the remaining cells. This method could provide a fairly accurate record of the distribution of the archaeological resource, provided the density of the sample points is high and evenly distributed across the landscape. Unfortunately, the Minnesota archaeological database is too sparse in many places and does not contain any direct, accurate measure of artifact density to make this approach feasible. Such an approach is most feasible for within-site studies (Kvamme 1990, p. 270).
Artifact diversity is a quantitative measure that could be included as an additional variable, along with artifact density, for estimating the archaeological significance of cells within a landscape. Together, density and diversity can measure the size and complexity of artifact assemblages per cell or site area. One advantage of this approach is that it makes no assumptions about the function of a site, the number of occupations at a site, or site age. This is significant, because it is just such kinds of information that are often absent or imprecise in the SHPO database. However, implementation of this would require the acceptance of standards for measuring or estimating density and diversity and additional work to add values for these fields to the SHPO database. Some preliminary work in this area was done in the early stages of this project. One shortcoming of this approach is that it may provide a false picture of site significance. Older, rarer sites are likely to have very low values for both of these measures.
Model stability is mapped in Figure 7.3 . A larger archaeological database, derived from probabilistic surveys, could enhance both the reliability of this measure and the stability of the models (Section 18.104.22.168.2). Assuming that a large, representative (probabilistic) sample of archaeological sites is available for modeling, model stability could be further improved by stratifying the sample into more homogeneous populations of sites.
At present, Mn/Model is essentially unstratified. Models predict the probability that archaeological resources of any type or time period are present or absent in a particular place. The problem with this type of model is that it best predicts the sites that are most abundant in the database. Older sites and special use site types are among the kinds of sites that would not be well predicted. Such sites would be considered statistical outliers in the database, in that the environmental characteristics with which they are associated may be very different than the majority of sites. As outliers, they contribute noise, or inconsistent information, to the database. By removing such noise from the database, the resulting sample is more homogeneous, therefore more easily predictable by statistical methods. Alternative methods for stratifying the models are discussed below.
The archaeological database includes sites extending over thousands of years. Minnesota’s environment has changed dramatically over that time. At historic contact, vegetation ranged from coniferous forest to prairie. Still earlier, the state was blanketed with a spruce parkland, then by prairie grasslands, and eventually by a mosaic of mixed hardwoods, deciduous forest, and prairie. Major rivers changed from fast-flowing discharge channels in the Late Glacial period to smaller, relatively placid modern rivers. Native Americans fired the grasslands to increase animal browse and, in so doing, changed the distribution of major plant communities.
All of these and other factors complicate the predictive modeling process. This potentially increases the percentage of unexplained variance (outliers) in site location by creating discordance between the virtual environment created by the GIS and past environments. It is expected that developing models for discrete time periods would provide better results. The geomorphological context described in Section 10.2.1.4 could be used to develop models of sites from distinct periods in the past with respect to contemporaneous environmental conditions.
Unfortunately, many sites in the state have not been assigned a date. This will limit the size of the sample that can be used to develop period-specific models. For regions with large numbers of known archaeological sites, a first step towards this approach may be to remove all paleoindian and archaic sites, as well as sites of unknown age, from the database before modeling. By modeling only Woodland and later sites, using the environmental data that are now available, the resulting models may provide more precise and more stable predictions of the most abundant sites in the database. Models for earlier periods will be dependent on the development of the geomorphological context for each period as well as the addition to the database of large numbers of archaeological sites from those periods.
Archaeological sites result from a number of different behavioral activities. Even within a discrete time period, these activities may be specific to certain locations. Kvamme (1988, pp. 381-385) discusses methods for developing models where site type is the dependent variable. An alternative would be to produce separate models for specific site types, such as settlements. However, site function is not always known. This reduces the effective sample size. Better site function data of this nature will improve future models. Even when function is known and sufficient numbers of sites are available, meaningful site-type categories must be defined (Kvamme 1988, p. 384). Poorly defined categories will almost certainly reduce the predictive power of the model.
Survey probability models were developed as part of this project to model the spatial biases of past surveys (Section 7.2.2). These models provide us with an approximation of the location and extent of environmental contexts that have been adequately surveyed. It follows that the existing survey data are representative of these parts of the landscape. This provides the opportunity to develop new models, with nonsites represented by negative survey rather than random points, that apply to these adequately surveyed areas only. Such models should be able to better articulate the locational differences between sites and nonsites within the kinds of environments that have been best studied.
The enhancements discussed in this section are aimed at improving the methods used to develop the present models. These methods are described in detail in Chapter 7. Suggestions for different modeling approaches are discussed in section 10.5.
Mn/Model was developed almost entirely on a UNIX-based GIS system. Since the beginning of this project, software products (ARC/INFO for NT, the ArcView Spatial Analyst extension and S-Plus for NT) have been developed that would allow the development of these models completely in a Windows NT environment. This has been successfully tested on several regions. Processor speed, available RAM, operating system stability, and hard drive capacity and access speed are the keys to efficiently using the NT platform to perform statewide, high-resolution modeling.
The advantages of the NT platform are several.
- The process can be made accessible to users and GIS installations that are not familiar with or equipped for UNIX.
- Costs for software and hardware are lower.
- Application development can be used to automate components of the modeling process and make these available to less sophisticated users via simple user interfaces. This can allow such users to not only consult models, but also perform analysis, maintenance or update functions.
Over the course of this project, two different regionalization schemes were used. The first, Archaeological Resource Regions, was replaced in Phase 3 by the Ecological Classification System (ECS, see Section 4.6). There are arguments for further reconsideration of the model regionalization scheme. Some of the arguments presented are discussed in this section.
The Minnesota River Valley should be modeled by itself. The valley per se, from bluff to bluff, is a very different ecological environment than the surrounding prairies. Moreover, since the valley is rich in sites, it may be one of the few, if not the only, similarly homogeneous units in the state with sufficient archaeological data to support a separate model. Furthermore, such a model could take advantage of the 1:24,000 scale landform sediment assemblage layer that was developed for the valley. Ideally, there will be enough sites remaining on the prairies in the Minnesota River Prairie Subsection, that they can be modeled separately as well. If this is the case, the results for both models should be superior to the current Minnesota River Prairie model, which represents more heterogeneous conditions.
The Red Wing area should be removed from the Blufflands subsection for modeling. The modern City of Red Wing occupies the site of a large prehistoric city. The large concentration of sites in and around Red Wing heavily weights the Blufflands archaeological database. This may mask patterns present in other parts of the region. These two portions of the Blufflands subsection could be separated, provided enough sites are available in each to support separate models. Again, unless site numbers are a problem, these models should be superior to the Blufflands model. One new model would provide articulation, within the very small but densely occupied Red Wing area, of where sites are more likely to be. The other would do a better job of articulating where sites should be away from Red Wing.
It has been implied that the ECS classification scheme is most appropriate for the relatively recent past (last 3,500 years) and may not do an adequate job of representing environments from 10,000 B.P. To 3500 B.P. If this indeed is the case, building models specifically for older sites (see Section 10.3.1) would require development of not simply a geomorphic context (Section 10.2.1.4), but also of a regionalization scheme for each time period of interest. This regionalization scheme could be based on the geomorphic context, as well as on climate model results and pollen data. Its development would require considerable analysis of the available data, so would be a considerable undertaking.
Finally, it has been suggested that perhaps the ECS system is not appropriate for archaeological predictive modeling because it was not developed for that purpose (Section 22.214.171.124). It may be true that a superior regionalization scheme can be developed, based on a thorough analysis of the archaeological and environmental data available. However, this would be a major research project. The risks would be high, since the cost of such an effort would demand that the resulting regions perform considerably better than the ECS subsections. When enough cultural data are available, a limited study of both the ECS and archaeological resource regions should be conducted to determine whether one system is superior, whether existing regional boundaries can be refined for modeling, or whether an entirely new regionalization scheme is needed.
Over the course of the project, there have been many suggestions for improving modeling procedures. Many of these were tried and evaluated. The best of these were incorporated into the development of the Phase 2 and 3 models (Chapters 4 and 7). However, time constraints prevented all of them from being tested. Some of these are discussed elsewhere in this chapter. Several more are mentioned here.
Several data layers used for modeling do not support the 30-meter resolution of the resulting models. Some of these data layers may be improved in the near future, while others may not. Two approaches have been suggested to improve the cartographic accuracy of the models. First, all low-resolution data layers could be excluded and models developed using on the highest resolution data. These models may be evaluated by comparison to the original models to determine whether the improved cartographic accuracy compensates adequately for the loss of information about vegetation, geomorphology, and other factors. Second, all data layers could be resampled to a larger cell size and models developed at that cell size. This would result in a loss of detailed (local) information from the high-resolution data layers, but may compensate for this loss with better performance in some regions.
Interpreting models can be difficult when they include large numbers of variables. However, the numbers of variables seem to be a function of the numbers of sites modeled, and models developed using more sites are more stable. Consequently, building models using fewer sites is not a viable option. It was suggested that to simplify such models, the variables from models in two adjacent regions be examined and that a subset of variables that are common to both models be used to build an alternative model. This would presumably increase interpretability and perhaps also reduce border effects. However, since this suggestion was made, the worst border effects (present in the Phase 2 models) have been corrected by the regionalization and modeling procedures used in Phase 3. Moreover, experience has shown that model performance may deteriorate if variables are removed, as that removes information about some site/environment relationships from the model. Perhaps a better solution would be to incorporate formal analytical methods into model interpretation to assess relationships between variables and allow the interpreter to evaluate the effects of related, correlated, or dependent groups of variables, rather than interpreting each individual variable. Moreover, graphic representation of these relationships (in 2-D or 3-D views) could be used to illustrate these relationships.
For every model, there are a certain number of sites that are not well predicted. These are the sites in the low probability zones. They may be considered outliers in the database, sites that are where they are for reasons very different from the sites the model best predicts. They may be older sites, or they may be functionally different from the majority (i.e. mound sites, where the majority of sites are camps). Just as it makes sense to examine these sites carefully to find patterns in how each model is not performing, it could also be informative to model these sites separately. However, the very low site numbers in the database may make this feasible for only a few subsections. Moreover, the resulting models may not perform well because these relatively small populations of sites may be quite diverse, consisting of sites that are outliers for a variety of different reasons.
The modeling procedures used assume that all sites are located with respect to necessary resources. This may be true for most kinds of sites used by hunter-gatherers in life, including villages, wild rice camps, hunting camps, quarry sites, and others. However, the dead do not require such resources, and burial sites may be located on the basis of very different, more culturally determined factors. It may be possible to improve the models by removing mounds and other burials which are likely to contribute "noise" to the database. Since the number of burial sites is relatively small, this should be possible in regions where site numbers are not extremely low.
The vegetation pattern at the time of the Public Land Survey had been in place perhaps only 600 years. Consequently, it has been argued that only the most recent sites should show strong relationships to the current vegetation variables. There are counter-arguments, primarily that local vegetation patterns are controlled largely by soils and terrain and that landscapes situations supporting the most mesic plant communities today probably supported the most mesic communities in the past, though the composition of these communities has probably changed. Still, it could be informative to run new models without vegetation variables, for comparison with the Phase 3 models. Moreover, if models are run for older sites, it is fairly clear that these vegetation variables should not be used.
More work could be done with modeling at a statewide level. The statewide model developed in Phase 3 of this project (Figure 8.8) serves to illustrate the relative probabilities of finding sites across the state. It is more accurate in this depiction than the composite of the regional models (Figure 8.5), because the regional models do not have comparable raw model values (Section 7.4.1). More work could be done interpreting this model. Moreover, it may be worthwhile to overlay this model with regional models to develop an adjusted model that may do a better job of portraying both regional and statewide site potential. Additional enhancements for understanding the statewide distribution of sites could include developing estimates of a priori probabilities for finding archaeological sites. This is done by dividing the sum of known site areas by the total area surveyed. Not all counties or regions have data that will support this kind of analysis. It has been done for the counties sampled in the 1996 Mn/Model survey. It may be possible to extend this analysis to some counties surveyed for Mn/SAS. Additional estimates of a priori probabilities may be possible once archaeological site and survey boundaries are mapped.
Models are in reality continuous probability surfaces, with each cell containing a unique value. However, to evaluate and compare models, it helps to minimize the number of values. Phase 1, 2, and 3 of the project each used different methods for classifying model scores into high, medium, and low probability values (Section 7.5.1). All three methods were arbitrary. The Phase 1 and 3 methods were more easily replicable than the Phase 2 methods. However, there would be some advantages to devising a well-founded, statistically based method for model classification. First, it may be possible to develop a scheme more grounded in statistical theory and therefore less arbitrary than the current method. Moreover, it could be made less confusing to apply and, consequently, more replicable. However, the advantage of the current scheme is that it is tied directly to project goals, though the goals themselves were arbitrarily set.
So far, models have been evaluated only for how well they predict known sites. Ideally, they should also be evaluated for how well they predict locations where sites are known not to be. This second test was somewhat stymied by the realization that negative survey points, the component of the database recording site absence, were distributed in much the same way as known sites. However, with the development of the survey models, it would make sense to further evaluate the Phase 3 models for only the portion of the landscape that qualifies as "adequately surveyed." This should be done using both site present and site absent locations. If models do not perform well on these tests, an argument could then be made to construct new models that apply only to the adequately surveyed portions of the landscape (Section 10.3.3).
Throughout this project, models have been evaluated on the basis of how well they met the stated goals (85% of sites in high and medium probability, which should constitute 33% or less of the land area). However, there has been no statistical test of the likelihood of the results being accurate. This evaluation could be improved by placing a confidence band around each prediction. The primary measure of model performance used so far has been Kvamme’s gain statistic (Kvamme 1988, p. 329). Other appropriate performance statistics should be identified and included in future evaluations.
A randomization procedure could be developed to test the probability of finding sites in each of the three mapped probability classes. This procedure would involving generating a vary large number of random points, determining how many of these fall into each probability class, and determining how many of these coincide with sites in each probability class. The percentage of points coinciding with sites in each probability class would provide an estimate of the probability of finding sites in that class.
Only one archaeological database has been available for both building and testing the models. In Phases 1 and 2, and for the preliminary models in Phase 3, this database was split into training and testing populations. However, if two sites were found in one survey, one site could be used to build the model and the other to test it. This violates the assumption that the training and testing data are independent. They clearly are not. New data, unrelated to the training database, provide the best model test. Testing the models with data from new field surveys as they become available will be part of MnDOT’s implementation plan.
As discussed in Chapter 7, the aim of predictive modeling in cultural resource management is to describe regularities in the locations and patterns of known archaeological remains to predict where other such remains may be found. This does not require the explanation of correlations between site locations and environmental variables. Some interpretation, or explanation, of the models has been provided in Chapter 8. However, this effort was limited by time constraints and project priorities. There are opportunities for other archaeologists to provide more in-depth interpretations of the models. Some approaches that have not yet been used for interpreting these models are mentioned here.
The primary aim of model interpretation is to explain why sites are located where they are. Some environmental variables associated by the models with site locations may not make sense in light of what is known about settlement patterns. These variables may be serving as proxies for other variables that are more directly related to the underlying human decision making processes, but cannot be measured. Posing and testing hypotheses about what these variables represent could improve our understanding of the models.
All of the model interpretations reported in Chapter 8 were carried out by the project archaeologist and Phase 3 research director. There is a great need for regional specialists to study and interpret each model with respect to their intimate knowledge of the region in question.
The model interpretations included in Chapter 8 were based on univariate and simple spatial analyses. Multivariate techniques, such as discriminate analysis and principal component analysis, could provide additional insight into the models.
In the discussion of the regional models (Chapter 8), hypotheses are suggested that could be tested to enhance the evaluation and interpretation of these models. However, some hypotheses have been posed that would involve consideration of multiple regions or comparisons between regions. These include:
- Except in the Blufflands and Rochester Plateau subsections, archaeological sites are near shorelines, and those that are not are usually small and occur at low densities.
- In areas where lakes are present, sites occur with greater density on lakeshores than on river or streamshores.
- In areas where lakes are present, 13-17.5% of surveyed 40-acre units in streamshore strata will contain sites (with the exception of the Blue Earth-Faribault area).
- In areas where lakes are present, only 0-8% of surveyed 40-acre units away from water will contain sites.
- In areas where lakes are present, sites will be concentrated in wooded areas of wild rice lakes at lake inlets and outlets.
- In areas with few lakes, sites occur in greater frequencies in streamshore strata than in streamshore strata in areas with many lakes, except in Blue Earth-Faribault.
- In areas with few lakes, site association with shoreline is usually less strong than in areas with many lakes.
- In areas with few lakes, the larger the differences in elevation above water, the greater the scatter of sites.
This project took a purely statistical approach to modeling the presence or absence of archaeological sites of all types (except single artifacts). This approach was chosen because it is scientific, replicable, and can efficiently handle very large volumes of data. However, it has its shortcomings. Foremost among these is that its performance relies on a sufficiently large database. Because the distribution of archaeological sites in Minnesota is quite sparse, there are few known sites in some regions. Since most alternatives to modeling site presence/absence require extracting subsets of the database (i.e. sites of a certain type), such models could not be built using statistical methods because of insufficient data.
Two alternative approaches are discussed here that may allow for modeling attributes other than site presence or absence.
Predictive models may be more useful for Cultural Resource Management if they focus on predicting significant archaeological resources. However, extracting only "significant" sites from the database for modeling results in low site numbers. An alternative would be to assign significance rating to each site in the database, then use this rating as the dependent variable in a linear regression model. The drawback to developing a model of site significance, however, is in the assignment of the ratings. First, the database must contain sufficient, accurate information for a very large number of sites to allow significance to be measured. Assuming this requirement is met, there must be clear criteria for determining which sites are significant and which are not.
The Florida Department of Transportation, with funds from FHWA, has been refining criteria for assessing "important" archaeological sites and building a model to predict the locations of such sites. Site significance ratings are based on interest to the public, potential scientific value, proximity to a similar site on protected land, the severity of threatened impacts, and other factors. The highest scores are assigned to the least well represented sites in the database, because excavations of such sites would produce more significant new knowledge. When the SHPO database has been enhanced with more complete and accurate data, it may be desirable to develop this kind of model for Minnesota.
Intuitive, or expert system, models are those based on archaeologists’ professional experience, rather than scientific methods. They may incorporate information from a variety of tangible and intangible sources, including archaeological theory, reports documenting site distribution patterns outside the region of interest, and field experience and knowledge gained from working with collections, other archaeologists, and the relevant literature within the region of interest. Their advantages lie in the incorporation of information that cannot be included in a database or analyzed statistically and the lack of reliance on large numbers of archaeological sites. Their primary drawback is that they are subjective, therefore not easily replicable by others. However, it should be possible to record many of the decision rules used in the model. These rules can then be applied to GIS data (provided the necessary information to support the rules is available) to map the model, or at least an approximation of the model.
Such non-statistical models provide an alternative for modeling site types that are not well represented in the database. Because they are not based on the existing archaeological database (which is biased towards more recent sites), intuitive models can use personal expertise to predict where sites poorly represented in the database might be found. This could allow for models focusing on "red flag" sites, i.e. those whose presence would cost projects money because of Phase 2 surveys, excavation, or alterations in construction plans (Altschul 1990). Intuitive models could be evaluated using the same methods applied to the statistical models and also by comparison to the statistical models.
Although large sites with dense and diverse artifact assemblages may be significant, most of these date to within the last 3500 years or so. Earlier sites are usually very small, but can be significant according to National Register criteria just because they are early and poorly known. These sites may be deeply buried or in unusual landscape positions today (e.g. because of isostatic rebound or the disappearance of a glacial lake).
One suggested approach to modeling such sites is to use the adaptive type concept, as outlined below. This approach recognizes the dynamic nature of Minnesota’s Late Glacial and Holocene environment. It requires the use of the paleoclimate/paleovegetation data developed for this project, as the geographical extent of specific adaptive settlement types expanded and contracted with major changes in paleoclimate. It would also make maximum use of the geomorphology data developed for this project.
The adaptive type approach is to build settlement models for predicting where sites of a particular settlement type are likely to occur. The remainder of this subsection outlines an interpretation of the basic adaptive types or human ecosystems that were present in precontact period Minnesota. By knowing where sites associated with particular human ecosystems, or adaptive types, are likely to occur; their absence in similar places becomes a CRM concern.
Adaptive types as used here refer to broadly defined patterns of subsistence. Examples are focal bison hunting, generalized hunting and gathering, and reliance on wild rice or maize as a staple. Since settlement pattern among hunter-gatherers is closely related to subsistence, the pattern of archaeological component locations associated with each subsistence method should be distinguishable from one another. Examples of published studies using a somewhat similar approach are Hammer (1993) and Stoltman and Baerreis (1983).
There are two main reasons for adopting an adaptive types approach in predictive modeling: (1) to help investigators more clearly understand why sites are located where they are; and, more importantly in CRM projects such as Mn/Model, (2) as a tool in identifying those types of components whose location is most difficult to predict (which will generally be Archaic and earlier components). The approach works by using statistical analysis to identify patterns of component locations that are distinguishable from one another, presumably because they are responses to different subsistence emphases. In addition, these patterns can be broken down during analysis into classes of site types for each adaptive pattern, such as warm weather / cold weather camps, wild rice harvesting / semi-permanent village / fishing / mortuary occupations, base camp / special extraction camp, etc.
Because gross subsistence-settlement types crosscut archaeological cultures, both unrelated and related archaeological cultures or traditions could be members of the same adaptive type. Since the same location could be seasonally occupied through time by people with different adaptive types, the same site could be a node in several different types of settlement patterns. It is more exact, then, to speak of component rather than site locations in these analyses.
In general, the conceptual boundaries between gross subsistence-settlement types are marked by major cultural and/or environmental discontinuities. Hammer (1993) suggests that these discontinuities reflect major changes in the conjunction of three related continua: (1) resource breadth (a focal to diffuse subsistence emphasis); (2) degree of seasonal mobility (from very mobile to sedentary throughout the year); and (3) complexity of social organization (level of cooperative effort). Seven adaptive types have been identified in this report on the basis of these criteria for precontact Minnesota: Late Glacial/Early Post-Glacial Hunters, Focal Bison Hunters, Generalized Hunter-Gatherers, Proto-Wild Rice Harvesters, proto-Horticulturalists, Intensive Wild Rice Harvesters, and Horticulturists. Almost all studies that employ an adaptive type approach find that the initial guesses as to basic adaptive types are wrong. Types useful for understanding and predicting site locations are the end product of extensive analyses. It should be no surprise, then, if the number of types, their definition, and the discriminators used to identify them change throughout this analysis.
The initial precontact adaptive site types identified for Minnesota are:
HG1. Late Glacial/Early Postglacial Hunter-Gatherers. Major change in environment. Highly mobile hunter-gatherers who made fluted projectile points. Includes Clovis (Cl), Folsom (Fo), and Eastern Fluted (EF) in the SHPO archaeological database. This mode disappears by 7000 B.C. It is more closely related to eastern fluted point adaptations in its more diffuse subsistence base compared to the Big Game Hunting adaptation on the Plains and in the Southwest (Gibbon and Johnson 1996).
FBH. Focal Bison Hunters. Major change in environment. Intensive bison hunting groups. Late Paleoindian, and Early and Middle Archaic, in all prairie regions west of a north-south line that runs through the Blue Earth River; all more recent groups outside of the Prairie Lakes region in the above area; the Middle Archaic within the area of expansion of the prairie to the northwest during the Prairie period. Includes all Prairie Archaic (AP) components in the SHPO archaeological database, except those dating to the Late Archaic period in the Prairie Lakes area. Also includes Lanceolate Point/Plano (Pl) in prairie areas.
HG2. Generalized Hunter-Gatherers. A general balance between meat and vegetable foods. Late Paleoindian through Middle Archaic in the southeast; Late Paleoindian through Initial Woodland (Malmo [HR in mixed hardwood forest], Brainerd [Br], Laurel [La] in the northern mixed-hardwood forests; Late Archaic, Fox Lake (FL), and Lake Benton (LB) in the Prairie Lakes region; to the contact period in far northern Minnesota outside the range of extensive wild rice beds. Includes Shield Archaic (AS), and any Lake Forest Archaic (AL) and Riverine Archaic (AR) components that date to the Middle Archaic period.
WR1. Proto-Wild Rice Harvesters. Although a trend occurred among some generalized hunter-gatherers toward increased reliance on plant foods (and perhaps especially wild rice), a real break does not seem to occur until after ca. A.D. 600 in the main wild rice growing areas of the northern mixed-hardwood forest. St. Croix (SO), Onamia (ON). Blackduck (Ba)-Kathio (Ka)-Clam River.
HR1. Proto-Horticultural Groups. Beginning with Late Archaic (many AR) in the southeast and extending up to the appearance of horticultural groups ca. A.D. 1000. Late Archaic in the southeast (many AR), Early Woodland (WL), Havana-related (Howard Lake and Sorg but excluding Malmo), Southeastern Minnesota Late Woodland (SELW), including all Madison Ware. Southeastern Minnesota is roughly defined here by a line that runs north south through the Blue Earth River and east west through St. Cloud.
WR2. Intensive Wild Rice Harvesters. Major change in social organization. Psinomani (Ps), which includes components associated with Sandy Lake and Ogechie ceramics (or that is labeled Wanikan). Applies only to the wild rice growing area of the northern mixed-hardwood forests.
HR2. Horticultural Groups. Oneota (O), Cambria (Ca), Great Oasis (GO), Big Stone (BS), and Silvernale (Sn). Oneota has been subdivided into Blue Earth (BE) and Orr (Or).
The Late Glacial/Early Postglacial Hunter-Gatherers type is separated from the onset of the Focal Bison Hunters and Generalized Hunter-Gatherer types by major changes in the environment (ca. 7000 B.C.), while the Intensive Wild Rice Harvesters and Horticultural Groups types are separated from earlier types by major changes in social organization (ca. A.D. 1000). The identification of subsistence-settlement types can be complicated by, among other things, gradual change through time in intensity of the defining characteristics of a type and by major shifts in the environment or subsistence emphasis that extend or contract the area in which a type is found. For example, there seems to have been a trend among some generalized hunter-gatherers toward increased use of plant foods, including wild rice. Nonetheless, a real break does not seem to occur in the main wild rice growing areas of the northern hardwood forests until ca. A.D. 600. Likewise, the mid-Holocene Prairie period greatly extended the area occupied by focal bison hunters (who are part of the Early and early Middle Archaic), just like the later inclusion of the Prairie Lakes region among proto-horticultural group shrank the area in Minnesota of focal bison hunters. Some types (Focal Bison Hunters, generalized Hunter-Gatherer) persisted in some areas of the state from Late Paleoindian times to the historic period, while others (Intensive Wild Rice Harvesters, Horticulturists) only first appear in the state sometime between ca. A.D. 1000-1250.
Finally, the majority of the precontact occupants of Minnesota seem to have wandered widely throughout the state and neighboring areas. For example, occupants of the northern hardwood forests typically traveled out onto the prairies seasonally to hunt bison. Since the focus here is on site location (rather than annual settlement pattern), the same group would act like a focal bison hunter in the prairies and a wild rice harvester in the forest. As a result, some discriminators for the presence of a type, such as a particular projectile point type, could indicate the presence of either adaptive type, depending on where the points are found. Unlike Hammer's (1993) example, therefore, in which a more homogeneous and stable environment was present throughout the Holocene, Minnesota's dynamic and diverse environment makes the use of discriminators less straightforward. Few widespread and reliable discriminators exist in precontact Minnesota sites for identifying adaptive types. The best available discriminators are projectile point and pottery styles, and, to a lesser extent, lithic raw material types. These also have problems that must be resolved for this type of analysis (e.g., the age and uses of particular point styles, the periods of popularity of lithic raw materials).
Since it is more reliable to identify contact period adaptive and site types through archival research, they are not discussed here, except to mention contact period culture types as listed in the SHPO archaeological database: Western Dakota (WD), Eastern Dakota (ED), Chiwere Siouan (CS), Ojibwa or Chippewa (Oj), French (Fr), English (En), and Initial U.S. (US).
Enhancements discussed so far in this chapter have focused on ways to improve the models themselves. However, another goal of great interest to MnDOT is enhancing the functionality, or usefulness, of Mn/Model. Such enhancements may involve extending the use of the models and environmental data to provide additional information for planning and resource management, integrating Mn/Model data with other data used in the organization, and extending the use of data developed for Mn/Model to contribute to different functions already performed within the organization or to support completely new applications. In short, Mn/Model's products will contribute to process improvement within MnDOT and, perhaps, in other state agencies and local governments. Three such enhancements are discussed below.
The first formal enhancement to Mn/Model was a prototype cost path analysis application for cultural resources and other environmental issues that are a factor in the NEPA process. The application, now available only as a prototype, allows planners and engineers to compare alternative corridor alignments and determine which has the lowest cultural resource and environmental mitigation costs. The analysis also generates a "least cost path" alignment based upon an analysis of the costs being modeled. The intent is to provide an advisory tool for environmental, land use, and transportation planners. Factors modeled in this decision support application include relative cost of archaeological survey, intrinsic value of wetlands, land value based on land use, and the relative costs of stream crossings.
Additional data can enhance not only the development of this application, but also the extension of functionality into other areas of impact evaluation. One data set that contributes greatly to the estimation of mitigation costs is Mn/DNR’s County Biological Survey, which identifies significant natural areas within the state, including the locations of rare and endangered species. Economic data, such as land ownership, land use, and right of way acquisition, could be factored into the analysis of project costs as the data become available. Other enhancements may also include adding data layers for such purposes as environmental justice (e.g. demographics, social analysis) and transportation planning (e.g. traffic counts, pavement management, accident data).
Phases 1 and 2 of Mn/Model were largely experimental. Many different data sets, variables, and modeling procedures were tested. Only in Phase 3 did "best practices" and a flexible database begin to emerge. MnDOT is currently in the process of developing formal process and data models for Mn/Model. The process model will focus on Phase 3 "best practices," but will also accommodate data and model update procedures as they are developed. The data model will consider the best Phase 3 data, the new higher resolution data sources more recently available, and the use of these data throughout M/DOT. Both models will increase the efficiency of the modeling process and software applications developed to use the models and data. They will also provide a context into which new data could be more readily integrated. Establishments of links between attribute tables and associated tables (such as the SHPO database, the NWI database, or the NRCS soils database) would facilitate the incorporation of new variables or new analyses in the future.
The formal process model and its accompanying user manual will be the culmination of the extensive documentation prepared for Mn/Model. It should facilitate knowledge transfer to other states wishing to adopt or adapt the methods used for developing Mn/Model.
This project, which is currently underway, is exploring the most efficient and cost-effective means of probing sediments that are likely to contain deeply buried sites. An array of methods are being tested on both known buried sites and locations indicated as having high potential for buried sites according to the geomorphic models. Both geophysical and geoarchaeological methods are being evaluated. The result will be a protocol for MnDOT to follow under an array of conditions when testing for buried sites.
Mn/Model is by no means a static product. The ideas suggested in this chapter for enhancing Mn/Model are by no means exhaustive. The data and models can provide the basis for process improvement, model enhancements, and archaeological research for many years to come.
With so many possible directions to follow, Mn/Model enhancements must be prioritized to provide the maximum possible benefit to Mn/Model for the cost and to meet the agency’s most immediate needs. For process improvement, this means integrating Mn/Model with other data to provide information that can save time and costs.
For Phase 4, the priority will be to make the best use of new environmental data that have become available. Development of better archaeological data will take more time, but in the long run may contribute more than anything to model improvement. Other priorities in Phase 4 include the integration of the geomorphic models with the statistical models and an attempt to model locations of drained lakes and abandoned river channels, at least for one region. These efforts will begin to incorporate information about past environments into the environmental variables.
Research conducted independently by archaeologists can provide valuable testing and interpretation of the models. However, for this to happen may first require convincing a skeptical archaeological community of the value of a GIS-based predictive model. Lack of confidence in this type of model is not unique to Minnesota. A national survey conducted for this project found several reasons for such skepticism. These are listed below, along with comments indicating how these concerns have been addressed by the Mn/Model implementation plan (Chapter 11):
- Many previous models have not worked in actual application (i.e. have been unreliable predictors of site location). Overcoming this obstacle will require demonstration of the reliability of the models through the kind of testing proposed in the implementation plan.
- Predictive models predict the locations of the types of sites that were used in the construction of the models and not all sites, such as very early sites (whereas intuitive models can take these concerns into consideration). This is a good argument for archaeologists to develop their own intuitive models into a formal set of mappable decision rules. The mapped results can then be compared to Mn/Model and formally evaluated. This will be required for individual project areas, under the Mn/Model implementation plan.
- Concern for their misuse, with the fear that models could be used as rationale for not surveying many areas that in reality contain sites. This concern is addressed by MnDOT’s implementation plan that provides for sampling both the areas predicted as high probability by Mn/Model and the areas identified as high probability by the archaeologists intuitive models. The plan also provides for sampling of low probability and unknown areas.
- Fear that use of predictive models will reduce the amount of survey (and thus money they earn). Again, the Mn/Model implementation plan makes it clear that survey will be required on each project, according to criteria that will provide for both protecting cultural resources and testing the effectiveness of the models. These survey strategies are expected not to reduce surveys, but rather to increase the value of the information gained from the surveys.
- Pride in their own intuitive models. Mn/Model’s implementation plan provides for the incorporation of archaeologists’ intuitive models into the survey strategy for the project. Comparing the effectiveness of the intuitive models and Mn/Model could lead to information that can be used to improve both.
The kinds of tests Mn/Model will be subjected to as part of the implementation plan are exactly what is required to guide future model improvements. If these tests do not empirically demonstrate that this statistical approach works as well as or better than intuitive models, then clearly the models must be improved or empirically validated intuitive models substituted. The CRM community needs highly successful models that can be applied by a wide variety of individuals.
1992 The Late Holocene History of Prairie, Brush-Prairie, and Jack Pine (Pinus banksiana) Forest on
Outwash Plains, North-Central Minnesota, USA. The Holocene2(1):37-50.
1993 A Groundwater Model to Explain Past Lake Levels at Parkers Prairie, Minnesota USA. The Holocene
Almquist-Jacobson, H., J.E. Almendinger,
and S. Hobbie
1992 Influence of Terrestrial Vegetation on Sediment-Forming Processes in Kettle Lakes of West-Central
Minnesota. Quaternary Research.
1990 Red Flag Models: The Use of Modelling in Management Contexts. In Interpreting Space: GIS and
Archaeology, edited by K.M.S. Allen, S.W. Green, and E. B.W. Zubrow, pp. 226-238. Taylor and
Francis, New York.
1995 Geological Contexts of the Early and Middle Holocene Archaeological Record in North Dakota and
Adjoining Areas of the Northern Plains. In Archaeological Geology of the Archaic Period in North
America, edited by E. A. Bettis, pp. 67-86. Special Paper 297, Geological Society of America, Boulder,
1996 Cultural Response or Geological Process? A Comment on Sheehan. Plains Anthropologist
Baker, R.G., L.J. Maher, C.A. Chumbley,
and K.L. Van Zant
1992 Patterns of Holocene Environmental Change in the Midwestern United States. Quaternary Research
Baker, R.G., D.P. Schwert, E.A.
Bettis III, and C.A. Chumbley
1993 Impact of Euro-American Settlement on a Riparian Landscape in Northeast Iowa, Midwestern USA:
An Integrated Approach Based on Historical Evidence, Floodplain Sediments, Fossil Pollen, Plant
Macrofossils and Insects. The Holocene 3(4):314-323.
Bartlein, P.J., and C. Whitlock
1993 Paleoclimatic Interpretation of the Elk Lake Pollen Record. In Elk Lake, Minnesota: Evidence for
Rapid Climate Change in the North-Central United States, edited by J. Platt Bradbury and Walter E.
Dean, pp. 275-293. Special Paper 276, Geological Society of America, U.S. Geological Survey, Denver.
Bartlein, P.J., and C. Whitlock
1993 Paleoclimatic Interpretation of the Elk Lake Pollen Record. In Elk Lake, Minnesota: Evidence for
Rapid Climate Change in the North-Central United States, edited by J.P. Bradbury and W. E. Dean,
pp. 275-293.Special Paper 276, Geological Society of America, U.S. Geological Survey, Denver.
Bell, J. and M. Krusemark
1998 Current Research at the U of M, Department of Soil, Water and Climate’s Soil and Landscape Analysis
Lab Regarding Soil Survey Mapping in Minnesota. GIS/LIS News 20 (Fall 1998):6.
1990 Depositional Stratigraphy, Site Context and Prehistoric Cultural Overview. Holocene Alluvial
Stratigraphy and Selected Aspects of the Quaternary History of Western Iowa, pp. 75-85. Guidebook
Series No. 9, Geological Survey Bureau, Iowa City.
Bettis, E.A., III (editor)
1995 Archaeological Geology of the Archaic Period in North America. Special Paper 297, Geological
Society of America, Boulder, Colorado.
Bettis, E.A., III
1990 Holocene Alluvial Stratigraphy of Western Iowa. Holocene Alluvial Stratigraphy and Selected
Aspects of the Quaternary History of Western Iowa, pp. 1-16. Guidebook Series No. 9, Geological
Survey Bureau, Iowa City.
Bettis, E.A., III., and D.W. Benn
1984 An Archaeological and Geomorphological Survey in the Central Des Moines River Valley, Iowa. Plains
Bettis, E.A., III, and E.R. Hajic
1995 Landscape Development and the Location of Evidence of Archaic Cultures in the Upper Midwest. In
Archaeological Geology of the Archaic Period in North America, edited by E. A. Bettis, pp. 87-113.
Special Paper 297, Geological Society of America, Boulder, Colorado.
1969 The Archaeology of Petaga Point: The Preceramic Component. Minnesota Prehistoric Archaeology
Series, No. 2. Minnesota Historical Society, St. Paul.
Boudreau, D. and E. Hobbs
1994 Historical Vegetation GIS from Land Survey Records. Paper presented at the Fourth Annual
Conference of the Minnesota GIS/LIS Consortium, St. Louis Park, MN.
Bradbury, JP, and W.E. Dean (editors)
1993a Elk Lake, Minnesota: Evidence for Rapid Climate Change in the North-Central United States.
Special Paper 276, Geological Society of America, U. S. Geological Survey, Denver.
1993b Holocene Climate and Limnologic History of the North-Central United States as Recorded in the
Varved Sediments of Elk Lake. In Elk Lake, Minnesota: Evidence for Rapid Climate Change in the
North-Central United States, edited by J. Bradbury and W. Dean, pp. 309-328. Special Paper 276,
Geological Society of America, U. S. Geological Survey, Denver.
Bradley, R.S., and P.D. Jones (editors)
1992 Climate Since AD 1500. Routledge, New York.
1983 Chronology of Lake Agassiz Drainage to Lake Superior. In Glacial Lake Agassiz, edited by J.T. Teller
and L. Clayton, pp. 291-307. Geological Association of Canada Special Paper 26. Department of
Geology, Memorial University of Newfoundland, St. John's.
Dean, W.E., T.S. Ahlbrandt, R.Y.
Anderson, and JP Bradbury
1996 Regional Aridity in North America During the Middle Holocene. The Holocene 6(2):145-155.
Drexler, C.W., W.R. Farrand, and
1983 Correlation of Glacial Lakes in the Superior Basin with Eastward Discharge Events from Lakes Agassiz.
In Glacial Lake Agassiz, edited by J. T. Teller and L. Clayton, pp. 309-330. Geological Association of
Canada Special Paper 26. Department of Geology, Memorial University of Newfoundland, St. John's.
Eyster-Smith, N.M., H.E. Wright,
Jr., and E J. Cushing
1991 Pollen Studies at Lake St. Croix, a River Lake on the Minnesota/Wisconsin Border, USA. The
Forester, R.M., L.D. DeLorme, and
1987 Mid-Holocene Climate in Northern Minnesota. Quaternary Research 28:263-272.
Fritz, S.C., D.R. Engstrom, and
1994 ‘Little Ice Age’ Aridity in the North American Great Plains: A High-Resolution Reconstruction of
Salinity Fluctuations from Devils Lake, North Dakota, USA. The Holocene 4(1):69-73.
Gibbon, G.E., and H. Hruby
1983 First-Step Settlement Subsistence Models for the Rock River Drainage in Southwestern Minnesota. In
Prairie Archaeology, edited by Guy Gibbon, pp. 131-150. Publications in Anthropology No. 3.
University of Minnesota, Minneapolis.
Gibbon, GE, and C. Johnson
1996 Modeling Paleoindian and Early Archaic Settlement in Minnesota. Paper presented at the 54th Annual
Plains Anthropological Conference, October 30 - November 2, Iowa City, Iowa.
1998 Status of Digital Soil Survey Mapping in Minnesota. GIS/LIS News 20 (Fall 1998):1, 3.
Grigal, G.F., R.C. Severson, and
1976 Evidence of Eolian Activity in North-Central Minnesota, 8,000 to 5,000 Years Ago. Bulletin 87, pp.
1251-1254. Geological Society of America, Denver.
1981 An Ecological and Paleoecological Study of the Vegetation in the Big Woods Region of
Minnesota. Unpublished Ph.D. dissertation, University of Minnesota, Minneapolis.
1988 The Little Ice Age. Methuen, New York.
1990a Late Pleistocene and Holocene Landscape Evolution, Depositional Subsystems, and
Stratigraphy in the Lower Illinois River valley and Adjacent Central Mississippi River. Unpublished
Ph.D. dissertation, Department of Geology, University of Illinois, Urbana.
1990b Koster Site Archaeology I: Stratigraphy and Landscape Evolution. Research Series Vol. 8.
Center for American Archaeology, Kampsville, Illinois. 1993. Geomorphology of the Northern American
Bottom as Context for Archaeology. Illinois Archaeology 5:54-65.
1993 Geomorphology of the Northern American Bottom as Context for Archaeology. Illinois Archaeology
1993 A New Predictive Site Location Model for Interior New York State. Man in the Northeast 45:39-76.
Holiday, V.T. (editor)
1992 Soils in Archaeology: Landscape Evolution and Human Occupation. Smithsonian Institution Press,
1980 Plant Geography, 2nd Edition. St. Martin’s Press, New York.
Keen, K.L., and L.C.K. Shane
1990 A Continuous Record of Eolian Activity and Vegetation Change at Lake Ann, East-Central Minnesota.
Bulletin 102, pp. 1646-1657. Geological Society of America, Denver.
1987 Historical Valley Floor Sedimentation in the Upper Mississippi Valley. Annals of the Association of
American Geographers 77:224-244.
1993 Large Increases in Flood Magnitude in Response to Modest Changes in Climate. Nature 361:430-432.
1988 Development and Testing of Quantitative Models. In Quantifying the Present and Predicting the
Past: Theory, Method, and Applications of Archaeological Predictive Modeling, edited by W.J. Judge
and L. Sebastian, pp. 325-428. U.S. Department of the Interior, Bureau of Land Management, Denver,
1990 The Fundamental Principles and Practice of Predictive Archaeological Modeling. In Mathematics and
Information Science in Archaeology: A Flexible Framework, edited by A. Voorrips, pp. 257-295.
Studies in Modern Archaeology, Vol. 3. Holos-Verlag, Bonn, Germany.
1998 Updating Outmoded Soil Surveys. GIS/LIS News 20 (Fall 1998):6-7.
Lasca, N.P., and J. Donahue (editors)
1990 Archaeological Geology of North America. Centennial Special Volume 4, Geological Society of
1992 Soils and Holocene Landscape Evolution in Central and Southwestern Kansas. In Soils in Archaeology,
edited by V. T. Holiday, pp. 41-117. Smithsonian Institution Press, Washington, D.C.
1995 Geomorphic Controls of the Archaic Record in the Central Plains of the United States. In
Archaeological Geology of the Archaic Period in North America, edited by E. A. Bettis, pp. 87-113.
Special Paper 297, Geological Society of America, Boulder, Colorado.
1974 The Original Vegetation of Minnesota. Compiled from U.S. General Land Office Survey notes. U.S.
Department of Agriculture, Forest Service, North Central Forest Experiment Station, St. Paul, Minnesota.
1983 River Warren, the Southern Outlet of Glacial Lake Agassiz. In Glacial Lake Agassiz, edited by J. T.
Teller and Lee Clayton, pp. 231-244. Geological Association of Canada Special Paper 26. Department of
Geology, Memorial University of Newfoundland, St. John’s.
1998 NRCS Soils Work in Minnesota. GIS/LIS News 20 (Fall 1998): 5.
McKinnon, N.A., and G.S.L. Stuart
1987 Man and the Mid-Holocene Climatic Optimum. Proceedings of the Seventeenth Chacmool
Conference, University of Calgary Archaeological Association, Calgary, Alberta.
1984 Discovering Sites Unseen. Advances in Archaeological Method and Theory 7:223-292.
1995 Modeling the Pleistocene Response to Altithermal Climates on the Southern High Plains. In Ancient
Peoples and Landscapes, edited by E. Johnson, pp. 349-368. Museum of Texas Tech University,
Mooers, H.D., and C.A. Dobbs
1993 Holocene Landscape Evolution and the Development of Models for Human Interaction with the
Environment: An Example from the Mississippi Headwaters Region. Geoarchaeology 8(6):475-492.
Needham, S., and M.G. Macklin (editors)
1992 Alluvial Archaeology in Britain. Monograph 27, Oxbow Books, Oxford.
1996 The King Coulee Site. Unpublished Masters Thesis, Interdisciplinary Archaeological Studies Program,
University of Minnesota.
1993 A Time-Space Model for the Distribution of Shoreline Archaeological Sites in the Lake Superior Basin.
1973 The Concept of an Altithermal Cultural Hiatus in Northern Plains Prehistory. American Anthropologist
Root, M.J., and S.A. Ahler
1987 Middle Holocene Occupation and Technological Change in the Knife River Primary Source Area. In
Man and the Mid-Holocene Climatic Optimum, edited by N. A. McKinnon and G. S. L. Stuart, pp. 85-
109. Proceedings of the Seventeenth Chacmool Conference, University of Calgary Archaeological
Association, Calgary, Alberta.
Running, G.L., IV
1995 Archaeological Geology of the Rustad Quarry (32RI775): An Early Archaic Site in Southeastern North
Dakota. Geoarchaeology 10:183-204.
1987 Formation Processes of the Archaeological Record. University of New Mexico Press, Albuquerque
1995 Cultural Responses to the Altithermal or Inadequate Sampling? Plains Anthropologist 40:261-270.
1995 Geoarchaeological Perspectives on Paleolandscapes and Regional Subsurface Archaeology. Journal of
Archaeological Method and Theory 2(1):69-104.
Stafford, C.R., D.S. Leigh, and
1992 Prehistoric Settlement and Landscape Change on Alluvial Fans in the Upper Mississippi River Valley.
1987 Deposits for Archaeologists. Advances in Archaeological Method and Theory 11:337-395.
Stein, J.K., and W.R. Farrand (editors)
1985 Archaeological Sediments in Context. Peopling of the Americas, Vol. 1, Center for the Study of Early
Man, Institute for Quaternary Studies, University of Maine, Orono.
Stoltman, J.B., and D.A. Baerreis
1983 The Evolution of Human Ecosystems in the Eastern United States. In Late Quaternary Environments
of the United States, vol. 2: The Holocene, edited by H.E. Wright, Jr., pp. 252-270. University of
Minnesota Press, Minneapolis.
Teller, J.T., and L.H. Thorleifson
1983 The Lake Agassiz-Lake Superior Connection. In Glacial Lake Agassiz, edited by J. T. Teller and Lee
Clayton, pp. 261-290. Geological Association of Canada Special Paper 26. Department of Geology,
Memorial University of Newfoundland, St. John's.
1964-1969 Composite Map of the United States Lane Surveyors Original Plats and Field Notes (for
Minnesota). Published by Trygg Land Company, Ely, Minnesota.
Van Nest, J.
1993 Geoarchaeology of Dissected Loess Uplands in Western Illinois. Geoarchaeology 8:281-311.
Warren, R.E., and D.L. Asch
1996 A Predictive Model of Archaeological Site Location in the Eastern Prairie Peninsula, Illinois. Unpublished
manuscript, Illinois State Museum, Springfield.
Webb, T., E.J. Cushing, and H.E.
1983 Holocene Changes in Vegetation of the Midwest. In Late Quaternary Environments of the United
States, edited by H. E. Wright, pp. 142-165. University of Minnesota Press, Minneapolis.
1976 The Dynamic Nature of Holocene Vegetation - A Problem in Paleoclimatology, Biogeography, and
Stratigraphic Nomenclature. Quaternary Research 6:581-596.
Wright, H.E., Jr., K. Lease, and
1995 Lake Pepin and the Environmental History of Southeastern Minnesota. Unpublished Manuscript.
The Mn/Model Final Report (Phases 1-3) is available on CD-ROM. Copies may be requested by e-mail: email@example.com
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.
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.