Reports

REGIONAL CLIMATIC ANALYSES OF

THE NORTHERN LAKE STATES

LAKE STATES NATIONAL FORESTS



Submitted By:

George E. Host and Philip L. Polzer
Natural Resources Research Institute
University of Minnesota, Duluth
5013 Miller Trunk Highway
Duluth, MN 55811

and

David T. Cleland
North Central Experiment Station
5985 Highway K
Rhinelander, WI 54501

NRRI Technical Report NRRI/TR-95/32
Center for Water and the Environment Contribution Number 162
NRRI Geographic Information Systems Number 34

The University of Minnesota is an equal opportunity educator and employer.

Introduction
Methods
Results
Summary
Citations

Introduction
Climate and regional physiography form the highest levels of multifactor ecological classification systems used within the Lake States region (Bailey 1985, Barnes 1982). While the glacial geology of the Lake States has been fairly well documented, climatic patterns have been assessed only at relatively coarse spatial scales. Rauscher (1984) used principal component analyses of weather station data to identify 20 macroclimatic regions within the Lake States. While the classification gives a good picture of regional climatic patterns, these classification units are generally too broad to be used for ecological classification purposes; often only one or two units will occur within a National Forest. Denton and Barnes (1987, 1988) conducted a climatic classification for the state of Michigan, which was subsequently used for identifying landscape ecosystems within the state (Albert et al. 1986). Even this finer-resolution level classification, however, had problems in areas where climatic changes were relatively continuous, as in the western Manistee National Forest (Albert et al. 1986). Little regional-scale climatic work has been conducted in Wisconsin and Minnesota.

In conducting Ecological Classification and Inventory (EC & I) on National Forests, as well as for regional-scale interagency planning efforts, there is a clear need for climatic maps of finer spatial resolution, on the order of 1:250,000 to 1:500,000. These maps should be able to distinguish variations in climate attributable to lake effects, cold air drainage, and physiographic variables. Recently, digital data on monthly precipitation and minimum and maximum temperatures on a 1 km² basis have become available. These commercial products use interpolation algorithms based on 30 year weather data to generate gridded data files (ZedX, Inc.). They represent a formidable amount of computer processing of information from the National Climatic Data Center, and have undergone extensive validation testing. Data are georeferenced to a grid of 30 second longitude x 30 second latitude, which is amenable for processing with raster-based geographic information systems.

The objective of this study was to use this high resolution climatic data to conduct a climatic analysis and classification of Michigan, Wisconsin, and Minnesota. Remote-sensing based analyses were used to identify fine-scale units; these were then aggregated into more homogeneous units based on similarity. The output products are maps and GIS coverages of the final classification in conjunction with associate data tables; special emphasis is placed on National Forest lands.

Methods
We identified climatic regions based on a high resolution climatic database consisting of temperature and precipitation values interpolated over a 1 km² grid across the study area (Russo et al. 1993). Source data were monthly averages of minimum and maximum temperatures and precipitation derived from 30-year (1961-1990) climatological station summaries published by the National Climatic Data Center. Multiple regressions were used to model climatic variables as a function of latitude, longitude and elevation; the model was then used to predict values over the 1 km² grid (Williams and Liebhold 1995). The accuracy of the interpolations was evaluated by comparing model predictions with actual station values. The average absolute differences between predicted and observed minimum and maximum temperatures were 0.56 and 0.72 C, respectively; the average absolute difference for precipitation was 0.70 cm (ZedX, Inc., 1995). The 36 climatic data layers (12 months x 3 data types) were incorporated into ARC/INFO, a polygon-based GIS (ESRI 1992), as point coverages, and subsequently converted into raster format using ERDAS image processing software (ERDAS 1991).

The 36 data layers were subjected to univariate and multivariate analysis to identify climatically homogeneous regions based on similarities in seasonal temperature and precipitation profiles. Principal component analyses (PCA) were conducted on each of the three data sets, and eigenvalues analyzed to identify those months that contributed most to the overall variance of the data set. These variables were carried on to further analyses. We also conducted PCA on a 15 layer data set using representative months stratified across the growing season: January, March, June, August, and October. We used five rather than twelve months to reduce the strong temporal autocorrelations between months (Briggs and Lemin 1992). These ordination procedures allowed us to identify variables important across the entire data set, and within each individual data set.

Variables identified in the PCA analysis were the inputs to a 15-band ERDAS iterative clustering algorithm. This analysis assigns a cluster for each pixel based on minimum spectral distance and then iteratively recalculates the means of each cluster until the means no longer shift (Tou and Gonzalez 1974, ERDAS 1992). The number of classes identified in this cluster analysis was user defined; we set an initial cluster size of 100, resulting in roughly an order of magnitude greater intensity than previous classifications. The output classes were then aggregated into larger, more homogeneous classes based on similarities within the dendrogram; 23 larger units were identified. Mean values of each class for each data set used in the final analysis were calculated and reported in tabular and map format.

Results

Principal Component Analysis

In the principal component analysis of 12 months of 30 year average minimum temperatures, 89% of the overall variance was loaded onto the first axis. As might be expected, winter temperatures showed the greatest degree of variation across the Lake States, with December, January, and February receiving loadings of 0.45, 0.45 and 0.39, respectively. On the second PCA axis, which accounted for 10% of the overall variance, May through July minima received high positive loadings. The distribution of February minimum temperatures across the Lake States is shown in Figure 1. The coldest temperatures are distributed along the northern border of Minnesota; the Superior National Forest encompasses much of this cold region. In northern Wisconsin and Michigan, Lake Superior moderates these colder temperatures. Only a small portion of the Ottawa NF receives this moderating effect, however, much of the forest is in the colder interior of the Upper Peninsula. The northern districts of the Chequamegon experience similar temperature effects, but the southern unit and much of the Nicolet receive some moderating effects due to latitude. Similarly, the Hiawatha has relatively warm winter temperatures due to the moderating effects of two lakes. The strong moderating effect of Lake Michigan is evident on the Manistee National Forest, where a gradual climatic gradient extends from the Lake to the interior highlands. The prevailing wind patterns, coupled with the large outwash surfaces in western Michigan result in a less steep climatic gradient compared with Lake Michigan's effect in eastern Wisconsin. The distribution of minimum temperatures is clearly one of the most important factors controlling the distribution of species in the Lake States.

Figure 1. Minimum February temperatures interpolated across Michigan, Wisconsin, and Minnesota.

A similar pattern was found with maximum temperatures. Variation in maximum temperatures is greatest during the winter months; December and January were heavily weighted on the first axis, which accounted for 85% of the overall variance. The second axis, accounting for 13% of total variance, was driven by May through August temperature maxima, with August receiving the highest eigenvalue. In contrast to minimum temperatures, the gradient of maximum temperatures has a strong SW-NE component, with hottest August temperatures in the continental interior (Figure 2). The prairie-forest transition of western Minnesota is well-defined by maximum temperatures, as is the 'tension zone' through the three states.

Figure 2. Maximum August temperatures interpolated across Michigan, Wisconsin, and Minnesota.

Seasonal variation in 30 year average precipitation was driven by the amounts of precipitation in November, December and March. The first axis accounted for 77% of the overall variation. November precipitation showed a strong gradation from NW to SE, ranging from 0.5 inches in northwestern Minnesota to 2.8 inches in SE Michigan (Figure 3). The strong elevational controls found in the temperature data were not evident in the patterns of precipitation. Precipitation in May, June and July were the most heavily weighted variables on the second PCA axis. This axis accounted for 13% of the overall variation, but is likely more biologically significant than the first axis.

Figure 3. November precipitation interpolated across Michigan, Wisconsin, and Minnesota.

Cluster Analysis

The full suite of variables selected from PCA as inputs to the classification are presented in Table 1. Similarities among pixels were calculated based on these variables, and an agglomerative classification performed. The study region was classified into 95 units as a function of this similarity matrix (5 of the 100 classification units specified in the original analysis had no pixels assigned to them and were dropped).

Table 1. Variables receiving extreme eigenvalues in each data set.

The high resolution classification integrates the numerous gradients defined above. This categorical representation of continuous data delineates the latitudinal gradient of minimum temperatures, the SW-NE gradient of maximum temperatures, and the SE-NW gradient of precipitation (Figure 4). Evident in this classification are the prairie-forest transition, the tension zone, and to some degree the southern boreal forest transition zone. Units range from relatively large, homogeneous units in the peatlands of northern Minnesota to smaller units in the highly dissected terrain of Wisconsin's driftless area. Since climatic models were developed independently for different states, there is some degree of discontinuity at state boundaries. These discontinuities, however, are within error bounds calculated for the data (J. Russo, personal communication).

Figure 4. High resolution classification of the Lake States based on PCA derived input variables within an isodata clustering routine.

For comparison with other classification systems, the high resolution classification was aggregated based on similarity among units. An agglomerative clustering routine was used to combine similar units. Logical cut levels were selected based on the aggregation patterns shown in the clustering dendrogram. Twenty-three aggregated climatic regions were identified (Figure 5).

Figure 5. Aggregation of high resolution classification into mesoclimatic regions for Michigan, Wisconsin, and Minnesota.

Summary
Climate is one of the main factors controlling both forest productivity and regional species range distributions. It is also a major hierarchical level in all EC & I systems in Region 9 (Cleland et al. 1992), as well as in analogous systems being developed at the state level. Yet, because of difficulties in synthesizing and analyzing information, it is one of the most poorly quantified of ECS data layers. A high-resolution climatic classification for the northern Lake States represents a valuable asset for specialists working in the development and interpretation of ecological units. In a related study, for example, we used this climatic database to develop a detailed climatic classification of the area in and around the Chequamegon National Forest (Host et al. 1996). Other studies have used this information to predict susceptibility of sites to insect defoliation (Williams and Liebhold 1995).

The ordinations and classifications described here are based on monthly thirty-year averages. As a result, there are several aspects of climate important to plant growth, such as growing season, that were not represented in the input data. This current classification thus represents climatic regions based on thirty years of record, and should be considered an initial high resolution stratification of the Lake States. More detailed analyses, incorporating AVHRR data, greenness indices, and other components of the weather record can be superimposed on this framework to conduct more specific, localized analyses.

Ecological classifications are the upper levels of the hierarchy are important for addressing research questions which extend beyond National Forest boundaries. Included in these are the role of the Forests in multi-agency regional landscape management, the assessment of biodiversity and ecological risk at regional scales, and the identification of ecologically-sensitive areas in the face of changing climatic conditions. Higher order classifications provide a context within which local variation in ecosystem properties may be assessed and quantified. This particular work supports the Lake States initiatives of subsectional classifications, off-Forest Landtype Association development, and the development of a framework for assessing the ecological health of Lake States forests.

Literature Cited

Albert, D. A., S. A. Denton, and B. V. Barnes. 1986. Regional landscape ecosystems of Michigan. School of Natural Resources, University of Michigan, Ann Arbor. 32 p.

Bailey, R. G. 1985. The factor of scale in ecosystem mapping. Envir. Manag. 9:271­276.

Barnes, B. V., K. S. Pregitzer, T. A. Spies, and V. H. Spooner. 1982. Ecological forest site classification. J. For. 80:493­498.

Briggs, R. D. and R. C. Lemin, Jr. 1992. Delineation of climatic regions in Maine. Canadian Journal of Forest Research 22:801­811.

Cleland, D. T., T. R. Crow, P. E. Avers, and J. R. Probst. 1992. Principles of land stratification for delineating ecosystems. Taking an ecological approach to management. USDA Forest Service Watershed and Air Management Publication WO-WSA-3:40-50.

Denton, S. R. and B. V. Barnes. 1987. Spatial distribution of ecologically applicable climatic statistics in Michigan. Can. J. of For. Res. 17:598­612.

Denton, S.R. and B.V. Barnes. 1988. An ecological climate classification of Michigan: a quantitative approach. For. Sci. 34:119-138.

ERDAS. 1992. ERDAS field guide. Version 7.5. ERDAS, Atlanta, Georgia, USA.

ESRI. 1992. Arc/Info User's Manual. Environmental Systems Research Institute, Redlands, California, USA.

Host, G. E., P. Polzer, D. J. Mladenoff, M. A. White, and T. R. Crow. 1996. A quantitative approach to developing regional ecosystem classifications. Ecological Applications 6:608-618.

Rauscher, H. M. 1984. Homogeneous macroclimatic zones of the Lake States. USDA Forest Service Research Paper NC­240 39 p.

Tou, J. T., and R. C. Gonzalez. 1974. Pattern recognition principles. Addison­Wesley, Reading, Massachusetts, USA.

Williams, D. W., and A. M. Leibhold. 1995. Forest defoliators and climate change: potential changes in spatial distribution of outbreaks of western spruce budworm (Lepidoptera: Tortricidae) and gypsy moth (Lepidoptera: Lymantriidae). Environmental Entomology 24:1­9.

ZedX. 1995. Database description: Minnesota, Michigan, and Wisconsin. Hi­Rez Data Climatological Series, ZedX, Boalsburg, Pennsylvania, USA.

Return to top