Comparisons of image- and plot-based estimates of number and size of forest patches in Michigan, USA

Mark D. Nelson1, Dacia M. Meneguzzo2, and Mark H. Hansen3

U.S. Department of Agriculture, Forest Service, Northern Research Station, Forest Inventory and Analysis, 1992 Folwell Avenue, St. Paul, MN, 55108, USA; 1-651-649-5104;

2 U.S. Department of Agriculture, Forest Service, Northern Research Station, Forest Inventory and Analysis, 1992 Folwell Avenue, St. Paul, MN, 55108, USA; 1-651-649-5129;

3 U.S. Department of Agriculture, Forest Service, Northern Research Station, Forest Inventory and Analysis, 1992 Folwell Avenue, St. Paul, MN, 55108, USA; 1-651-649-5148;

Introduction

National Forest Inventories (NFIs) typically use sample plot data to estimate metrics of forest vegetation composition and structure. In addition to their traditional use for timber assessments, NFI estimates can be useful for quantifying abundance of wildlife habitat, potential spread of non-native invasive species, and other ecological processes. However, habitat suitability and other ecological characteristics depend not only upon forest composition and structure, but also on landscape patterns (Wenny et al. 1993). Metrics used to quantify these patterns, such as number and size of forest patches, usually are derived from satellite image-based data, in the form of land cover maps (McGarigal and Marks 1995). Field plot data also have potential uses for estimating some metrics of landscape pattern. Plot designs used by many NFIs allow for either a cluster-plot (Kleinn 2000), or mapped-plot (Van Deusen 2005) approach, or both, to estimating metrics of landscape pattern. The U.S. Department of Agriculture, Forest Service, Forest Inventory and Analysis (FIA) program, which is the NFI of the United States, employs a plot design that provides for both approaches.

Different methodologies for estimating landscape patterns often yield different estimates. Nelson et al. (In Press) reported per-county estimates derived from Michigan, USA FIA subplot observations and a cluster-plot approach as having larger mean patch size and fewer numbers of patches than for estimates derived from an enhanced version of the 1992 National Land Cover Dataset (NLCD). Similarly, Meneguzzo (2008) reported that a mapped-plot approach with FIA plot observations resulted in estimates of mean forest patch size substantially different from NLCD 2001-based estimates, for three study areas in Michigan. However, comparison of independently derived image-based and plot-based estimates of landscape metrics often are confounded by several factors, including definitional differences of forest land use and land cover, temporal inconsistencies, classification inaccuracies, spatial mis-registration between plot location coordinates and image pixels, indistinct ecotones between forest and nonforest land, sensitivity to sample size, and scale effects. To assess the utility of plot-based estimation approaches while controlling for these confounding effects, we (1) produced a satellite image-based geospatial dataset of forest land cover patches, (2) intersected locations of FIA subplot center points and circular polygons with forest patch polygons to simulate FIA subplot- and condition-level observations, (3) produced estimates of number and mean size of forest patches by applying both cluster-plot (Kleinn 2000) and mapped-plot (Van Deusen 2005) approaches to the simulated plot observations, and (4) compared the simulated cluster-plot and mapped-plot estimates with each other, and with the land cover map-based estimates.

Study Area, Data, and Methods

We selected three study areas (SA) in Michigan, USA: SA1 - the northern Michigan county of Marquette, comprising 4714 km2 (excluding open water), of which 88 percent is forested; SA2 - the three adjacent central Michigan counties of Alpena, Montmorency, and Presque Isle, comprising 4616 km2 (excluding open water), of which 70 percent is forested; and SA3 - the three adjacent southern Michigan counties of Berrien, Cass, and Van Buren, comprising 4336 km2 (excluding open water), of which 32 percent is forested (Figure 1).

Figure 1. Forest (green) and nonforest (tan) land cover for the state of Michigan, USA, derived from NLCD 2001, with three study areas delineated.

FIA Plot Data

FIA defines forest land as being at least 10-percent stocked by trees of any size, not subject to nonforest use, at least 0.4047 ha in area, and at least 36.58 m in width. FIA ground plots consist of a cluster of four circular subplots - one central and three peripheral - each 7.32 m in radius. At each subplot, the center point is assigned a land use class, all land use boundaries are mapped, and trees are measured. Peripheral subplot centers are spaced 36.58 m from the center of the central subplot, at azimuths of 0, 120, and 240 degrees, forming an equilateral triangle with a circumcircle 73.15 m in diameter through the subplot centers. The FIA standard, base federal sampling intensity is one plot per approximately 2400 ha. See Bechtold and Patterson (2005) for a more detailed description of FIA plot and sample designs. During Michigan’s first annual inventory, the sample intensity of FIA plots was three times greater than the base federal sample intensity. Actual FIA plot observations were not used in this study. Rather, simulated observations were obtained with a geographic information system (GIS) by overlaying locations of subplot center points and circular polygons on an NLCD 2001-derived polygon map of forest patches, described below.

We simulated plot-based estimates of landscape metrics using cluster-plot (Kleinn 2000) and mapped-plot (Van Deusen 2005) approaches. Both approaches utilize plot dimensions and configuration to determine the probability of a randomly located plot intersecting a patch boundary based on assumptions of patch shape. A circular forest patch shape was assumed. Three peripheral subplot points and the single central subplot polygon were utilized for observations for the cluster-plot and mapped-plot approaches, respectively. A conservative test of statistical difference was conducted by comparing each of the simulated plot-based estimates of mean patch size to the 95 percent confidence interval surrounding NLCD 2001-based mean patch sizes.

Satellite Image-Based Land Cover Data

NLCD 2001 was used to obtain polygon delineations of forest patches for each of the three study areas. See Homer et al. (2007) for a detailed description of NLCD 2001. Polygons were derived with GIS software, as follows. NLCD classes were recoded into ‘forest’ and ‘nonforest’ land classes. Adjacent pixels of the same class were grouped together to create clusters with a minimum of four 30-m pixels (0.36 ha minimum); disjunct individual pixels and clusters of fewer than four like pixels were recoded to the class of the surrounding pixels. Pixel clusters were identified as patches and assigned a unique numeric code. The raster patch dataset for each study area was exported to a polygon layer in shapefile format. Nonforest patches were excluded from further analysis.

GIS software was used to produce a second dataset of patches, consisting of circular, non-overlapping polygons centered over each forest patch. The constraint that circular polygons be non-overlapping resulted in a substantial reduction in mean patch size for SA1 and SA2, where larger patches predominated, so the circular polygon dataset was analyzed only for SA3, where patches were smaller and more separated. First, the geometric centroid was determined for each forest patch, and was not constrained to fall within the original patch polygon. Second, the minimum geographic distance was computed between each centroid and its nearest neighbor. Third, a circular radius was calculated for each centroid, based on half the minimum distance between centroids, minus 0.5 m (to prevent circle overlap). Fourth, one circular polygon surrounding each forest patch centroid was created. Summary statistics were computed in a GIS to obtain mean patch size and number of forest patches for all three study areas, and for circular polygons within SA3.

Results

Mean size of forest patches from NLCD 2001 were 229.1 (n=1835), 51.0 (n=6422), and 11.7 ha (n=11912), for SA1, SA2, and SA3, respectively (Fig. 2). Estimates based on simulated plot condition observations using the mapped-plot approach were 190.6, 21.4, and 5.7 ha, for SA1, SA2, and SA3, respectively. Estimates based on the cluster-plot approach and simulated subplot observations, were 290.1, 58.6, and 9.2 ha for SA1, SA2, and SA3, respectively. Figure 3 portrays estimates of number of forest patches, which are equivalent to total area of forest land divided by mean patch size. Using the mapped-plot approach, simulated plot-based estimates of mean forest patch size in SA3 were 51.7 percent smaller than for NLCD 2001 forest patch polygons and are 0.1 percent smaller than circular polygons. Using the cluster-plot approach, simulated plot-based estimates of mean forest patch size in SA3 were 21.6 percent smaller than for NLCD 2001 forest patch polygons, and 34.2 percent larger than for circular polygons.

Figure 2. Mean size (ha) of forest patches for three study areas in Michigan, USA, based on the 2001 National Land Cover Dataset (NLCD) and simulated plot observations derived from the NLCD 2001 dataset, using mapped-plot and cluster-plot estimation approaches. Error bars represent ± 95 percent confidence intervals surrounding NLCD 2001 mean patch size.

Figure 3. Number of forest patches for three study areas in Michigan, USA, based on the 2001 National Land Cover Dataset (NLCD) and simulated plot observations derived from the NLCD 2001 dataset, using mapped-plot and cluster-plot estimation approaches.


Discussion

It appears that estimates from the simulated mapped-plot approach significantly underestimated mean patch size for SA2 and SA3, and the estimate for the simulated cluster-plot approach significantly underestimated mean patch size for SA3, when forest patch shapes were assumed to be circular.Compared with results for NLCD 2001 forest patches, which are irregular in shape, use of circular polygons appeared to eliminate the deviation between plot- and pixel-based estimates for the cluster-plot approach, but increased the deviation between estimates for the mapped-plot approach.

FIA cluster plots contain four subplots, but results reported here are for estimates produced when using only the central subplot for the mapped-plot approach, and the three peripheral subplots for the cluster-plot approach. Additional investigations are needed to more fully utilize all four FIA subplots, to assess variance estimators for the two plot-based approaches, and to determine the relative merits of both the mapped-plot and cluster-plot approaches to estimating landscape metrics.

Literature Cited

Bechtold, W.A., and Patterson, P.L. (eds). 2005. The enhanced Forest Inventory and Analysis Program - national sampling design and estimation procedures. U.S. Department of Agriculture Forest Service, Asheville, NC.

Homer, C., Dewitz, J., Fry, J., Coan, M., Hossain, N., Larson, C., Herold, N., McKerrow, A., VanDriel, J.N., and Wickham, J. 2007. Completion of the 2001 National Land Cover Database for the Conterminous United States. Photogrammetric Engineering & Remote Sensing 73(4): 337-341.

Kleinn, C. 2000. Estimating metrics of forest spatial pattern from large area forest inventory cluster samples. Forest Science 46(4): 548-557.

McGarigal, K., and Marks, B.J. 1995. FRAGSTATS: Spatial pattern analysis program for quantifying landscape structure

USDA Forest Service, Pacific Northwest Research Station. Gen. Tech. Rep. PNW-GTR-351.

Meneguzzo, D.M. 2008. Quantifying and monitoring forest fragmentation using satellite imagery and forest inventory and analysis plot data. In College of Food, Agricultural and Natural Resource Science. University of Minnesota, St. Paul, MN. p. 98.

Nelson, M.D., Hansen, M.H., and Lister, A.J. In Press. Estimating number and size of forest patches from FIA plot data. In Proceedings of the Eighth Annual Forest Inventory and Analysis Symposium, Monterey, CA.

Van Deusen, P.C. 2005. Mapped plot patch size estimates. In Proceedings of the Fifth Annual Forest Inventory and Analysis Symposium. Edited by Ronald E. McRoberts, Gregory A. Reams, Paul C. Van Duesen, and William H. McWilliams. USDA Forest Service, Washington, D.C., New Orleans, LA. pp. 111-116.

Wenny, D.G., Clawson, R.L., Faaborg, J., and Sheriff, S.L. 1993. Population density, habitat selection and minimum area requirements of three forest-interior warblers in Central Missouri

The Condor 95: 968-979.

5