Multi-spatial scale representation of landscape transitions using landsat thematic mapper data and scale-space filters

Graduation date: 2003This thesis considered current approaches to describing landscape pattern,\ud identified scale issues associated with defining objects, and explored techniques to\ud reliably group elements based on land cover as represented by satellite imagery. It\ud was recognized that there is an important need to develop tools that can be applied\ud using remotely sensed data to objectively identify landscape features at multiple\ud spatial scales. The thesis focused on the potential for the use of specific digital\ud image processing techniques to objectively generate multi-scale landscape\ud heterogeneity models from Landsat Thematic Mapper (TM) satellite imagery. An\ud approach using Difference-of-Gaussian (DoG) scale-space filters was used to\ud process TM satellite data, transformed into the Normalized Difference Vegetation\ud Index (NDVI), to define discrete landscape units across a range of scales.\ud A series of Difference-of-Gaussian filters, with standard deviation (a) pairs of 0.9\ud and 1.1, 0.8 and 1.2, and 0.7 and 1.3, was applied to TM data. Results were\ud compared against published ecoregion classifications and aerial orthophotography.\ud These comparisons showed that the applied filters failed to accurately define\ud landscape elements at a range of spatial scales. Filter size was found to be highly\ud sensitive to the resolution of the image and object boundaries were maldelineated\ud when complex boundary combinations were encountered. A modified approach using Gaussian/Difference-of-Gaussian or\ud GaussianlLaplacian filter combinations and an alternate approach of watershed\ud segmentation were recommended as potential techniques to test in further study.\ud Future study should also include statistical approaches to aggregate zones and\ud relate delineated zones to an actual, on-the-ground physiognomic classification\ud scheme