Mn/Model

Elevation data, a key environmental layer, were inconsistent across the state (Figure 4.2). First, the 7.5 minute (1:24,000 scale) digital elevation models (DEMs) were not completed for the entire state at the beginning of this project. Consequently, 1:250,000 scale DEMs had to be used for parts of the state. Second, the 7.5 minute DEMs were of varying quality. One group of these, the Level 1 DEMs, was built to meet an accuracy standard set by USGS (U.S. Geological Survey 1990). Elevations were collected by the USGS and several independent contractors using various methods. One method produced distortion in the data, which resulted in a banding or striping effect when the data are viewed. This pattern was apparent in some Level 1 DEMs for Minnesota, mostly in the southeastern part of the state (Figure 4.2). Various algorithms can be used to reduce the systematic distortions (see Appendix B). However, because the distortion itself varies, no single algorithm removes it equally well from all DEMs where it occurs. By changing and standardizing methods for building DEMS, USGS eliminated this problem from the Level 2 DEMs. USGS and the State of Minnesota have now cooperatively completed Level 2 DEM coverage of the state, including replacing all Level 1 DEMs.

4.5.1.4 Completeness

GIS data that have not been completed for the area of interest or that are lacking attributes needed for analysis are poorly suited to a project. Several data layers that were considered very desirable for inclusion in Mn/Model were completed for only parts of the state at the beginning of the project. In the case of the 1:24,000 elevation data, lower quality (1:250,000 scale) data were substituted only for quadrangles that were not yet available at 1:24,000 scale. However, this introduced more inconsistency into the Mn/Model elevation data (discussed in Section 4.5.1.3).

Where it was not possible to complete coverage of the state using an alternative data source, data with incomplete coverage could not be used for statewide models (see Section 4.5.2). To assess the value of some of the incomplete data for future modeling, small area models were developed in Phase 2 using variables derived from county soil surveys and digitized Trygg (1964-1969) maps in addition to the variables used for the statewide modeling (Chapter 8).

Other components of data completeness include whether the data have been completely developed or processed to include all of the features needed for analysis and whether these features have been completely attributed. The MnDOT BaseMap includes a layer of lakes and double line rivers digitized from USGS 1:24,000 scale quadrangles. Coverage is statewide. However, the features exist only as line work (i.e. shorelines are represented). For GIS analysis, lakes and double line rivers must be represented as polygons and the polygons must contain attributes that distinguish lakes, islands, rivers, and land from one another. These polygons can be built from the existing lines using additional data processing steps. This is very time consuming, as it involved making sure that all lines are connected (in this case they were not) and adding a label point to each polygon for attaching attributes. Consequently, these data were not yet at a development stage where they could be used for modeling.

Complete attribute data are also important. Desirable attributes may be missing altogether. As an example, the available 1:24,000 scale coverages for lakes and rivers (from NWI) and perennial and intermittent streams (from the MnDOT BaseMap) were not attributed with the names of these features. Alternatively, attributes defined by a data provider may not be the ones needed for a predictive modeling project. For example, when the Mn/DNR 1:100,000 scale statewide geomorphology coverage became available late in the project, it was discovered that the attributes were not readily suited to the derivation of variables of interest to Mn/Model.

4.5.2 Consequences of Data Quality

There are a number of possible consequences of using poor quality data. Using multiple layers of data of varying quality can compound these consequences. The end result is that the quality of the final product, in this case the models, is a function of the composite quality of all of the data layers that went into building it. Some of the specific outcomes of using the lower quality data described above are discussed in this section.

Variables derived from low resolution databases may exhibit coarse spatial patterns (Chapter 8). This is an artifact of the data, which cannot make a distinction between one 30 meter cell and another. If the data were originally rasterized by 100 meter cells, for instance, regridding to 30 meter cells would produce clusters of nine cells containing the same value. If the variable is then used to build a model, the model may display clusters of nine cells with the same probability class. Sometimes these coarse patterns are overwritten in the models by the patterns contributed by variables derived from high resolution data. However, where they are apparent, the user should be aware that they are artifacts of the low resolution of the data and that the model could probably be improved (i.e. have greater spatial resolution) by better quality data.

In the course of this study, considerable bias and error in the available archaeological database has been documented (Chapters 5, 7, and 8). The older archaeological surveys are a particular problem. Because these data were not collected using probabilistic methods, they violate the assumptions of the statistical procedures and limit statistical confidence in the models. In Phases 1 and 2, sites from older surveys and modern surveys that did not meet specific criteria were not used for building models. However, even the carefully selected surveys used for Phase 1 and 2 modeling were demonstrated to be non-probabilistic, based on the distribution of their negative survey points (Chapters 7 and 8). In addition, problems of spatial accuracy were documented during Phase 3 (Chapters 5 and 12). Inaccurate attributes, such as the misclassification of site types or ages, must be assumed to be present as well. Moreover, the number of these qualified sites was very low, particularly in some regions (Chapters 7 and 8). All of these factors have the potential to weaken the predictive power of the models. However, the effects of low site numbers were most apparent in the preliminary models. To reduce these effects, and given that the other data problems documented made these sites less distinctive from the rest, the entire database was used to build the Phase 3 models. Knowing the extent of bias in the data, models of survey bias (survey probability models) were developed in Phase 3 to supplement interpretation of the site probability models.

The implications of the inconsistency of available elevation data in Minnesota are several. As always, data quality is relative to the applications for which the data are being used. Because USGS standards for representation of elevation are met, measurements of elevation and differences in elevation from all 7.5 minute DEMs (Levels 1 and 2) should be accurate at that scale. However, the elevations measured from 1:250,000 scale DEMs are more generalized, therefore less accurate when used at a 1:24,000 scale, as they were in Mn/Model. The quality of Level 1 DEMs that exhibit distortions is poor for the purposes of generating contour maps (Figure B.1) and modeling surface hydrology (Grabrecht and Starks 1995). It is likely that slope, aspect, and measures of local relief derived from these DEMs are also distorted, though there is no way to test this without high quality DEMs from the same area for comparison. Moreover, these same measures will be more generalized when derived from the 1:250,000 DEMs, with the effect that fine scale (1:24,000) relief features are not represented (Figure B.2). It must be assumed, for Mn/Model, that the accuracy of the derived variables surface roughness, slope and prevailing orientation varies across the state.

Data with incomplete coverage may be difficult or impossible to incorporate into a model. In a grid, a distinction is made between the value NODATA and the value "0". Whereas NODATA indicates that data are missing, "0" may mean that the elevation is at sea level, that there is no elevation difference, that the slope is perfectly flat, or that a cell is on the shoreline (0 m from the edge of a lake). Grid calculations that access an input cell with the NODATA value will return an output cell with the same value, even if the same cell in other input grids have valid values. This produces variable grids with areas of NODATA, which would not be accepted by S-Plus for modeling. Grid calculations will be misleading if a "0" or other artificial value is inserted where no data are available. This produces inaccurate variable grids and misleading models. Consequently, the best solution is to not use incomplete data for modeling.

Small area models developed in Phase 2 to evaluate the contributions of incomplete data layers were not completely successful (Chapter 8). First, for some of the areas modeled, the number of available sites was too low to produce a strong model. Second, the reduced environmental variability of small areas may limit the identification of relationships between sites and variables that might be detectable over a larger area. These results bode poorly for the future development of statistical predictive models using the 1:24,000 scale landform sediment assemblage data mapped for Mn/Model (Chapter 12). At the present time, these coverages map a very small percentage of the state’s land area. They include long, narrow river valleys that may cross several distinct environmental zones, as well as several isolated upland quadrangles. To develop these detailed maps for a large enough contiguous area within an environmentally homogenous region to make statistical predictive modeling worthwhile will take years.

Missing attributes may limit the variables that can be derived or the procedures that can be implemented. Because elevation data included elevations of the water surface rather than depths below the surface, it was not possibly to model the potential for finding sites within lakes. Because lakes, rivers and streams were not attributed with their names, the features could not be automatically labeled on the maps produced for this project. Nor is it possible to perform a query to count, for example, the number of sites within 200 meters of Pelican Lake. Moreover, the missing attributes slowed down documenting environmental variations and reporting of model results because digital map products (displaying variables and models) had to be cross-referenced to paper maps to determine the names of prominent features for reference in the written report.

Even when the desired attributes can be derived from the attributes provided with a database, the process can be very time consuming. The project geomorphologists worked with the attribute data from the 1:100,000 MN DNR geomorphology coverage during Phase 3 to incorporate attributes that would be consistent with their landform sediment assemblage maps and project goals. However, the new attributes were ready only after Phase 3 modeling was complete.

4.5.3 Data Confidence Model

In Phases 2 and 3, data confidence models were developed to illustrate the geographic distribution of data quality. Since this project necessitated the use of data that were more generalized than desired, flawed, or inconsistent across the state, data quality may vary not only between models but also within the extent of a single model. Understanding the relative quality of the input data is important for evaluating how much confidence to place in a model. Consequently these data confidence models become important tools for both implementation and for prioritizing future model improvements.

In Phase 2, a very simple algorithm was used to develop the data confidence model. It consisted of the addition of three values: the number of archaeological sites used to build the model, a code for the quality of the elevation data, and a code for the scale of the lowest resolution database that figured into the best model developed for that region (Table 4.3).

Table 4.3. Values for Variables in Phase 2 Data Confidence Model.

The Phase 2 algorithm places the greatest weight on data scale, the second greatest weight on the quality of elevation data, and the least importance on the number of sites used for modeling.

In Phase 3, variables and weights were reconsidered by the project team. The three variables used in the Phase 2 data confidence model were retained, and an additional variable, site density per square km by subsection, was included as well. While site numbers have a strong effect on the performance of models developed, one must also consider the size of the region modeled. Low site numbers, relative to other regions, may be an indication of the region’s size. Low site densities are likely to produce more variable, hence less reliable, model results.

The number of archaeological sites used to build the model and site density per square km were both mapped by ECS subsections (Figure 3.11). The sites used to build the model excluded single artifacts, as these were not used for building the Phase 3 models. Site density was calculated using all known sites, including single artifacts. The raw values for these variables are shown in Table 4.4. The lowest resolution database scale was mapped by ECS subsections (Figure 4.7). The scales of the lowest resolution data are shown on Table 4.4. The quality of elevation data was mapped by 7.5 minute quadrangle (Figure 4.2).

Table 4.4. Raw Values of Three Data Quality Variables in Phase 3 Data Confidence Model, by ECS Subsection.

To build the data confidence model, the values of the four variables were first normalized to a scale from 0-100. Site numbers and site density were normalized by the following equation:

Normalized Value = (Value – Minimum Value)/0.01(Maximum Value– Minimum Value)

The quality of elevation data and the lowest resolution database scale were normalized according to the values in Table 4.5.

Table 4.5. Normalized Values for Elevation Quality and Data Resolution Variables in Phase 3 Data Confidence Model.

The normalized values for each variable were assigned weights based on their assumed importance in determining model outcomes (Table 4.6). Numbers of sites and site densities were assigned the highest weights because of their demonstrated effects on model performance and variability (Chapter 7). Elevation quality and resolution have noticeable effects on the spatial patterns of the models and probably also on the variables selected. However, without higher quality data for the same region for comparison, it is impossible to evaluate their effect on model performance. Consequently they were assigned slightly lower weights.

The algorithm for determining data confidence values was:

Data Confidence = å _1-n(Normalized variable value * Variable weight)

The mapping units of the four variables figure largely in the geographic pattern of data quality (Figure 4.1). The effects of elevation data quality in enhancing or detracting from the overall data confidence in a region create patterns with straight line borders. The smallest of these represent individual 7.5 minute quadrangles. The largest represent groups of quadrangles of the same data quality. The effects of site numbers and site densities are apparent by the large patches lacking straight line borders. The resolution of the lowest quality data set used had little influence on the model, as 1:500,000 was the variable value for all but one of the modeled regions. The resultant model shows the highest data confidence in the Border Lakes subsection along the Canadian border, most of the Blufflands subsection along the southeastern border with Wisconsin, and part of the Big Woods subsection in south central Minnesota. All three of these regions have both high site numbers, high site densities, and unbanded 1:24,000 elevation data. A portion of the Minnesota River Prairie subsection in southwestern Minnesota also has comparably high data confidence values. This area has very high site numbers, but low site densities.

Table 4.6. Weights Assigned to Variables in Phase 3 Data Confidence Model.

4.6 REGIONALIZATION

In the context of this project, regionalization refers to the practice of dividing the state into a number of smaller, relatively homogeneous areas for the purposes of data development and modeling. This was desirable for several reasons.

• First and foremost, a state the size of Minnesota has a wide range of environmental variation. Factors associated with precontact site selection are expected to differ between forests and prairies or glaciated and unglaciated landscapes. It is assumed that some of these environmental regions in general are more attractive to, or were more inhabitable for, hunter-gatherers than others (Chapter 3). Models developed statewide would likely focus on variables with greater variation statewide than locally, losing information about locally important variables.
• Furthermore, in Minnesota, known archaeological sites are not distributed evenly throughout the state (Figure 4.3). In many cases, concentrations of known sites correspond to concentrations of surveys (Figure 4.4). At the same time, large areas of the state, particularly in the north and northeast, have not been surveyed. This would undoubtedly bias the high probability areas towards regions with more surveys and more known sites.
• Finally, models developed without regionalization would be likely to perform poorly, as the generalized site/environment relationships discerned would not adequately represent the specific situation in any part of the state. Either many sites would not be predicted by the model, or a very large land area would be classified as high/medium probability to predict 85 percent of known sites.

Previous research has involved modeling areas much smaller than the state of Minnesota. Table 4.7 reports the study area sizes for 11 predictive modeling projects that preceded Mn/Model. The largest of these projects (Carmichael 1990) modeled approximately 11 percent of the area covered by Mn/Model. None of these researchers considered regionalization of their models. Consequently, there was not an available precedent for determining the best scheme for dividing Minnesota into regions for the purpose of developing Mn/Model.

Table 4.7. Comparison of Mn/Model Study Area Size with that of Selected Previous Predictive Models.

4.6.1 Phase 1 Regionalization

Initially, consideration was given to the development of regions based on like sets of environmental variables. However, to derive such a classification scheme would be a major research project in itself. For practical reasons, previously defined regions suitable for the model were sought. MnDOT Transportation District Boundaries were considered, but rejected because they were not based on either environmental or archaeological criteria. Attention turned to the nine archaeological resource regions (see Figure 3.10) and their subregions defined for the state of Minnesota by Anfinson (1990). The characteristics of these regions, their derivation, and reasons for selecting them for this project are described in Section 3.4. These nine regions were used for the initial regionalization of the models in Phases 1 and 2.

The regions are:

1.        Border Lakes
2.        Central Lakes Coniferous
3.        Central Lakes Deciduous
4.        Lakes Superior
5.        Northern Bog
6.        Prairie Lakes
7.        Red River Valley
8.        Southeast Riverine
9.        Southwest Riverine

 

In Phase 1, five regions and, within them, only the counties with probabilistic surveys were modeled (Figure 4.5). The regions modeled were Central Lakes Coniferous, Central Lakes Deciduous, Prairie Lakes, Southwest Riverine, and Southeast Riverine. The remaining regions (Red River Valley, Northern Bog, Lake Superior, and Border Lakes) were not modeled in this phase because they did not contain counties with probabilistic surveys.

Each Phase 1 region contained between 41 and 190 archaeological sites from probabilistic surveys that were used to create models. However, it was observed that models based on small samples, covering small areas (41 in Southwest Riverine, 31 in Nicollet County) performed as well as or better than models that used larger samples in larger areas. This implies that increasing the size of the region has a stronger negative effect on model results than reducing the sample size and area. Presumably, this is because smaller areas are more environmentally homogeneous.

4.6.2 Phase 2 Regionalization

When all 87 counties were modeled in Phase 2, some large regions were divided into their subregions (Figure 4.6) to increase the amount of environmental homogeneity within each area modeled and to reduce the digital data to more manageable file sizes. All 21 subregions defined by Anfinson (1990) would have been used for modeling had site numbers not been a problem. Consequently, modeling in Phase 2 was organized by the groupings of subregions indicated in the second column of Table 4.8. Where enough sites were present, subregions were modeled separately. Where too few sites were available for modeling, adjacent subregions were combined. In all, 15 models were developed.

Table 4.8. Archaeological Subregions and Subregion Groups Modeled in Phase 2.

Two problems were noted in the Phase 2 models that were thought to be possibly attributable to the regionalization scheme used. One of these was an edge effect, where linear boundaries of the Southeast Riverine and Northern Bog regions were strongly reflected in the distribution of model probability classes (Figure 8.4). Had these regional boundaries been based on natural features, such as rivers or ridges, the anomalies may have been less noticeable. However, it was immediately apparent that these straight-line patterns of high probability cells were not related to any natural features that would have influenced the locations of archaeological sites. Nor were they artifacts of the environmental data, such as boundaries between elevation data of different scales.

The second problem related to the regionalization scheme was that many of the archaeological resource regions and subregions contain a relatively wide range of environmental variability. The best models in Phase 2, as in Phase 1, came from regions with little environmental variability. The issue was not necessarily the size of the region modeled, but whether it overlapped into two or more geomorphic regions or vegetation zones when compared to the environmental data.

4.6.3 Phase 3 Regionalization

Towards the end of Phase 2, the Ecological Classification System (ECS) sections and subsections (Figure 4.7) became available in GIS format for Minnesota (Hanson and Hargrave, 1996). The derivation of this classification system is detailed in Section 3.5. Since it was developed to identify, characterize, and delineate units of land with similar climatic, geological, physical, and biological features, it was hypothesized that its use as a regionalization scheme for Mn/Model Phase 3 would provide regions with greater environmental homogeneity than was previously available. Moreover, since ECS boundaries are based entirely on natural features, it was hoped that the edge effect observed in Phase 2 could be reduced or eliminated by its use.

This regionalization scheme had several other advantages for Mn/Model.

• It is an integrated multi-factor classification scheme, based on climate, geology, geomorphology, and historic vegetation. As such, it is conceptually compatible with the multivariate approach used for modeling.
• It was developed by a team of scientists using a comprehensive, documented interdisciplinary approach. Consequently, it is more likely to be replicable than is the archaeological resource region scheme that was developed by an individual.
• It is conducive to modeling at multiple scales. In the future, modeling can be extended to lower levels of the hierarchy (smaller than Subsections).
• Because it is being developed nationwide, its use provides the potential for consistency across states and between agencies.
• Regional and subregional boundaries are based on natural features, primarily geomorphic features that have been in the same place since the end of the glacial period. The archaeological resource regions and subregions are largely delineated by modern political boundaries or relatively arbitrary straight lines.
• The ECS boundaries were defined at a higher spatial resolution than were the archaeological resource region boundaries.
• The ECS subsections were much more homogeneous with respect to our independent variables than were the archaeological regions, thus reducing "noise" in the analysis.
• Finally, it was designed to facilitate the understanding of relationships between single environmental components within homogeneous environmental regions. This is consistent with Mn/Model’s modeling approach.

For these reasons, ECS sections and subsections were adopted as the regionalization scheme for the Phase 3 models. This decision has been criticized locally by proponents of the Archaeological Regions used in Phases 1 and 2. Their argument is that there is archaeological meaning to that classification scheme, whereas archaeological data were not considered in development of the ECS. However, Mn/Model is by necessity an environmentally deterministic model. Too little is known about Minnesota's prehistoric cultures to incorporate cultural data into the modeling scheme. Since the only alternative is to predict site locations based on the characteristic environments in which human activity was concentrated, it is important that the regions modeled be environmentally homogeneous to the greatest extent possible. Heterogeneous environments otherwise create uninterpretable noise in the models and may have attributed to the edge effects observed in Phase 2.

4.6.3.1 Concerns Regarding the Application of ECS Regions to Mn/Model

Concern was expressed that the ECS regions were developed for natural resource management, not archaeological applications, and were therefore more representative of recent environments than might be appropriate for this project. However, the archaeological resource regions are also based on recent environments, emphasizing the characteristics of the Late Holocene and historic periods (Section 3.4). The Mn/Model team felt the ECS regions were most appropriate for modeling because of the number of environmental factors considered, the multi-disciplinary methods used for their delineation, the correspondence with significant geomorphic discontinuities, the resolution and quality of their source data, and their environmental homogeneity.

The greatest flaw with the use of modern data for regionalization is that climate zones and biotic provinces have shifted dramatically across Minnesota in the last 10,000 years (Figure 3.1). Geomorphic and topographic features change relatively slowly. Although erosion and deposition can transform local landscapes very quickly, the landforms created by Pleistocene glacial advances and retreats still dominate the Minnesota landscape (Figure 3.9). However, over the last 10,000 years these landforms have existed in constantly changing climates and biotic provinces (Figures 3.1, 3.4, 3.5, 3.6, and 3.8).

However, the contribution of time-transcendent topographic and geomorphic features to the local microclimate and patterns of vegetation distribution should not be underestimated. Areas historically protected from fire by virtue of their relief or their situation with respect to water bodies would likely have had the same kind of protection in the more distant past (Grimm, 1981). While these areas may have supported stands of Big Woods in the contact period, they likely supported oaks or other more drought resistant trees or shrubs in the drier prairie period. Consequently, they may have held the same kind of attractions for human settlement, fire protection and a source of fuel, throughout time. This does not, however, negate the importance of climate change's effects on settlement pattern on a larger scale, in the movement of people to different regions in search of wildlife or other resources and the very important local effects on lake levels and shifting stream courses. In addition, drier periods may have reduced site density and increased the probability of artifacts being buried by wind or water borne sediment.

To the limited extent that ECS regions rely on historic vegetation boundaries for their definition, they suffer from being time-static. However, the vague definitions of the archaeological resource region boundaries do not necessarily impart time-transcendence. Ideally, there should be different regionalization schemes for several time periods or cultural complexes in Minnesota's past. The potential for creating such regionalization schemes is discussed in Section 4.6.4.

4.6.3.2 Mitigating Factors

However, even the modern distribution of ecosystems within the landscape can provide useful information about past environmental patterns. This is because of the intimate nature of the relationship between vegetation and soils. The formation of distinctive soil types has traditionally been considered a function of climate, organisms (especially vegetation), parent material, topography, and time (Brady 1974:303; Eyre 1968: 25-44; Jenny 1980: 9-14, 160-197). Jenny (1980: 201-203) takes a different perspective, viewing soils and organisms as the ecosystem, which evolves together. He considers climate, topography, parent material, and time to be "state factors," which may act as controls on the evolving ecosystem, but are independent of it. These state factors may or may not be independent of each other. Yet his examination of the evolution of several ecosystems emphasizes both the great lengths of time that may be involved and the importance of variations in parent material (Jenny 1980: 207-245). Moreover, he determines that soil properties evolve slowly compared to the potentially rapid rates of vegetation change (Jenny 1980: 243). This relatively slow rate of soil development can be attributed to the time required for processes to change the nature of parent material and relief.

Where erosion and deposition are not prevalent, parent material and topography can be constant over long periods of time. These play a key role in determining microclimate and the relative moisture content of soils within a local landscape (Bannister 1976: 18-20; Jenny 1980: 39-41, 246-304; Scott 1974: 51-55). Although vegetation is an important contributor to microclimate and moisture relations, parent material or topography may limit the establishment of some kinds of vegetation, which are otherwise climatically adapted, in parts of the landscape (Cooper 1961; Etherington 1975: 154-216; Mooney 1974; D. Scott 1974).

Moreover, topography may control microclimate in ways that produce profound constraints on vegetation and soils. Propensity for flooding, for instance, will preclude the growth of vegetation that cannot tolerate saturated soils. In Minnesota, Grimm (1981, 1984) concluded that fire was the primary factor controlling the distribution of mesic deciduous forest, oak savanna, and prairie within the Big Woods region. Water bodies and topography strongly controlled fire patterns. He observed that "sites with identical physical characteristics but slightly different geographical locations, such as on the opposite sites of a river, can support very different types of persistent vegetation" (Grimm 1981: 176).

It is likely that such strongly controlling factors are to a certain extent time-transgressive. The Big Woods region was dominated by prairie during the mid-Holocene, was invaded by oak woodland beginning about 5,000 B.P., and was rapidly invaded by mesic hardwoods only about 300 years ago (Grimm 1983). These changes were most likely induced by increased precipitation and possibly decreased temperatures. Undoubtedly, however, fire has played a role in the region throughout at least the last 5,000 years. While oak savanna and prairie now persist only in the most fire-prone parts of the landscape, it is likely that the most protected places served as refuges for oaks and mesic tree species when the climate regime was drier and prairies were dominant. Historic vegetation patterns can therefore hold the key to past patterns and may prove useful to reconstructions of paleolandscapes.

4.6.4 Potential for Regionalization Based on Past Environments

Although an archaeological predictive model could clearly be improved by the use of a regionalization scheme based on information about earlier environments, there are obstacles to such an implementation.

• To support high resolution modeling statewide, the regionalization scheme must define at least 15-20 distinct regions. The development of such a scheme would require the availability of information about earlier environments at a scale of 1:500,000 or better. Paleoclimate and palynological data have not been developed at that scale in Minnesota. An evaluation of available palynological and paleoclimate data can be found in Section 6.3. Neither was deemed to provide the resolution needed to make them useful for Mn/Model.
• There is no single past environment. Incorporating specific paleoclimatic information would require definitions of not only regions but also of time periods. Such an approach would necessitate modeling only sites of a specified age or cultural affinity. This greatly reduces the number of sites available for modeling. Moreover, since this information for many sites in Minnesota is not known, site numbers would be reduced even further. Given the very low numbers of Archaic and Paleo-Indian sites in Minnesota, it would not be possible to model these statistically.

Most known archaeological sites in Minnesota date from the last few hundred years. Models based on ECS subsections should be appropriate for the majority of these sites. They are expected, however, to do a poor job predicting the older, but less abundant sites in the state. However, even if higher resolution paleoclimate and paleoecological data were available for modeling these sites, the low site numbers would prevent the development of these models using statistical methods. Given the constraints of site numbers, only one time-generalized model is now possible for Minnesota.

Return to Top

4.7 PROCEDURAL DESIGN

Not only is it important to design the physical aspects of the GIS system, it is critical to design the procedures that will be used for data development, maintenance, storage, security, analysis, and presentation. In a large project such as Mn/Model, design and documentation of all procedures provides for consistent results, products that will be suitable for subsequent tasks, and continuity in the face of staff changes. Subsequent automation of procedures, to the extent possible, provides further insurance of consistency. This does not imply, however, that these procedures should be written in stone. Provisions for correcting and updating procedures on the basis of experience, lessons learned, and changing project requirements must be part of the design.

For large GIS projects, it is customary to start with a pilot. This involves collecting the necessary project data for a very small area, developing procedures for the entire range of project tasks, and testing procedures on the pilot area. This has a number of advantages:

• It allows the evaluation of a wide range of data for the project. Data sets that are determined to be unsuitable can be discarded with little loss in investment.
• Since small geographic areas process more quickly, a pilot is an opportunity to evaluate a number of techniques in a relatively short time.
• A mistake made on a pilot county is a learning experience. A mistake made on the entire state is a considerable amount of time lost.

Careful selection of the pilot area is important. For Mn/Model, the pilot area selected was Nicollet County. It had several advantages:

• Databases that were not yet available statewide at the beginning of the project were available for Nicollet County. These included National Wetlands Inventory, 7.5 minute DEMs, and high-resolution soils.
• There were a relatively large number of known archaeological sites in the county, including sites from a probabilistic survey, which would allow development of prototype models.
• The county included a portion of the Minnesota River Valley, the first valley mapped by the project geomorphologists.
• It is a fairly small county, compared to others in Minnesota. A small pilot area size reduces data storage and processing requirements.

There are, however, some lessons that can’t be learned in the pilot phase.

• Experience on a pilot may not be a good basis for estimating the time required for modeling the entire state. The time required for some GRID functions increases linearly with the number of cells processed; for other functions, the time increases exponentially.
• Not all potential pitfalls will be discovered in the pilot. Expect more and bigger pitfalls when the project is underway.
• The pilot area may not include all of the data problems that will be encountered for the state as a whole. Procedures that worked smoothly in the pilot may have fatal flaws when applied to other geographic areas.
• Pilots may not provide the information needed to refine modeling techniques for large areas. Models derived for relatively small areas with large numbers of sites tend to perform well. When the same procedures are used for larger, more diverse areas, particularly those with low site numbers, unanticipated problems arise.

4.7.1 GIS Data Development

GIS data development includes both data conversion and data integration. Data conversion refers to the transformation of data from paper to digital format or from one digital format to another. Integration refers to processing necessary to allow different GIS data sets to work together. A distinction is made between vertical integration, which registers GIS layers for accurate overlay, and horizontal integration, which involves combining adjacent GIS datasets of the same theme so that the resulting digital map is topologically consistent (features are correctly connected at the adjacent map edges).

Data conversion and integration were the most time-consuming tasks for Mn/Model, as is common in most GIS projects, even though few datasets had to be digitized from paper maps. Conversion of data from EPPL7 GIS format to the ARC/INFO GRID format required for Mn/Model was involved because there was no direct translator available. For the majority of the data, which arrived in digital form, considerable processing was necessary to integrate it properly. Several processing tasks were commonly required for each layer (see Appendix B and the Mn/Model metadata). In addition, the processing of many large databases required careful management of file sizes and computer resources and formal quality control procedures.

4.7.1.1 Data Conversion

Data conversion is the most basic GIS procedure. Developing the models using ARC/INFO GRID software required that all data be in ARC/INFO coverage or grid format. The types of data conversion required for this project are summarized below.

Paper to Digital Conversion

Some geographic data were not available in digital format. Several layers that were considered potentially important were digitized from paper maps. These included Trygg maps, bedrock geology for some southeastern Minnesota counties, and the Mn/Model geomorphology project field maps. None of these layers was completed for the entire state because of time constraints. These layers could be used for modeling only in limited areas. Fortunately, none were considered critical for model development. Because digitizing takes many hours, it is not realistic to begin a statewide modeling project until statewide coverage of all critical layers is available digitally.

For Mn/Model, all digitizing was done using AutoCAD. The AutoCAD drawings were converted to PC ARC/INFO coverages (with a y-shift to preserve precision), then to ARC Export files, using ArcCAD software. From the ARC Export files, ARC/INFO coverages, then grids, were created. Digitizing procedures are described in more detail in Appendix B.

Generating Digital Maps From Coordinates

Point coverages can be generated from x and y coordinates contained in a database file. Where locations of data observations have been recorded as latitude/longitude or UTM coordinates read from topographic maps, this is useful procedure for creating a spatial database. The original database file can be joined to the resulting point coverage attribute table.

Coverages of the locations of archaeological sites and random points were generated from x/y-coordinates (UTM) recorded in digital database files received from SHPO (see Section 5.5 and Appendix B). Most of these coordinates were in NAD27. The coverages created were projected to NAD83, then converted to grids. Projection of the coverage does not change the values of the UTM coordinates in the attribute tables, which retain the original NAD27 coordinates.

Climate and pollen data were received as ASCII files with coordinates of weather stations and pollen cores in latitude/longitude decimal degrees. Point coverages were generated from these files, then projected to UTM NAD83. Climate and pollen surfaces were interpolated from the coverages in GRID (see Section 6.3 and Appendix B).

Digital to Digital Conversion

Most of the digital data received were in ARC/INFO coverages (vector format) and had to be converted to ARC/INFO grids (raster format) for variable derivation and modeling. However, the MGC100 database and many county soils databases came in EPPL7 raster format. These had to be converted to ERDAS images using EPPL7 software from LMIC (see Appendix B). The ERDAS images could then be converted to ARC/INFO GRID format.

4.7.1.2 Horizontal Data Integration

For this project, horizontal integration consisted of taking the data in whatever tiles it was provided and manipulating it so that it was tiled by the regions used for variable derivation and modeling. In Phase 2, variables were derived for the same 21 regions that were used for modeling. In Phase 3, in the interest of saving time, ECS subsections were combined into nine very large regions for the purpose of deriving variables. For Phase 3 modeling, the variable grids for each ECS subsection were extracted from these nine regions as needed.

Throughout this project, counties were used as the most basic tiling unit for two reasons. First, counties are smaller than potential modeling regions. They can be assembled into a variety of combinations to create regions. This allows flexibility in the definition of regions and allowed Mn/Model to easily move from archaeological resource regions in Phases 1 and 2 to ECS subsections in Phase 3. Second, MnDOT specified that final delivery of all data and models be tiled by counties to facilitate distribution via the MnDOT GIS server. Tiling the data by counties in the beginning saved work later in the project when deliverables were created.

Some data received were tiled by units smaller than counties. This required appending the data to form the county coverages or grids. National Wetlands Inventory and USGS DEMs were both distributed as one digital file per USGS 7.5 minute quadrangle. There are 1,745 quads in Minnesota. These were consolidated to form 87 county files. Likewise, soils data received as township files were joined to make counties.

Other digital data layers were received as statewide maps, in either raster or vector format. These included all of the MGC100 layers, the Marschner map, the DNR geomorphology coverage, tree species distributions, and bearing trees. These layers were split into regions for variable derivation and into counties for final distribution. Performing two splits (regional and county) is much more efficient than splitting by counties, then appending counties into regions.

4.7.1.3 Vertical Data Integration

All maps in the GIS must be carefully standardized to overlay properly. Most commonly, this involves map projection (see Section 4.3.1 ). Digital maps are distributed in a variety of projections, units, and datum. To make these data compatible with the standards established for the project sometimes required projecting from other map projects or coordinate systems to UTM. Projection was also needed to convert data already in UTM coordinates from NAD27 to NAD83 or to convert units from feet to meters.