Fractal pattern analysis derived from image segmentation of classified thematic mapper imagery: A multi-scale regionalization in Maine and North Dakota, United States of America

The search for better methods to apply satellite data in the analysis of ecological concerns requires a linkage between image regions and critical ground features. This study develops and tests a new multi-pass image segmentation algorithm using Thematic Mapper data from the North Dakota and Maine. The model extracts spectral and spatial image variation and uses this information to subdivide images into regions of homogeneity. These regions, not individual pixels, are used for classification and spatial analysis.; A successful model of polygon delineation should link polygons to landscape patches and patch boundaries to ecotones. Methods developed in the paper to evaluate raster-polygon structure demonstrate that polygons resulting from this model maintain large internal spectral homogeneity and exhibit greater spectral distance from adjoining polygons. The polygon-based classification in North Dakota resulted in an land-cover accuracy of about 80%, a 20% improvement over previous classifications that used less well defined polygons. The preservation of landscape pattern albedo is optimized, enabling better testing of hypotheses concerning the link between the landscape structure and ecological processes.; A similar algorithm for polygon delineation and classification was applied to satellite images for Maine. The fractal dimension was calculated for multiple landcover types over a large number of randomly placed windows, and interpolated to estimate the spatial variation of fractal dimension across the state. Principal component images were calculated and then clustered to produce groups defining fractal-based regions. The regions were combined into larger zones based on their similarity.; The delineated regions for each study area showed significant differences in relative proportions of land-cover type. Boundaries showed significant correspondence with previously developed regionalizations, based on environmental parameters. In Maine, the regions are best described by land use, topography, and climate; in North Dakota, the regions are best described by land form and land use. The methods of this research provide a link between a fine-scale landscape structure and broad-scale ecological pattern. The application of the methods within two vastly different ecosystems imply a portability across a wide variety of landscapes.