Portland State University. Department of Geography
Date of Award
Master of Science (M.S.) in Geography
1 online resource (viii, 70 p.) : ill.
Landscapes -- Remote sensing, Geographic information systems, Spatial ecology
The effects of scale influence all aspects of spatial analysis and should be expressly considered early in research planning. Remotely sensed images provide unique landscape perspectives and possess several features amenable to dealing with scale. In particular, images can be segmented into image objects representative of landscape features and structured as nested hierarchies for evaluating landscape patterns across a range of scales. The objectives of this research are to evaluate methods for: 1) characterizing candidate image objects to inform the selection of user-supplied segmentation parameters and 2) exploring the multi-scale structure of landscape patterns for defining and describing potentially important scales for conducting subsequent geospatial and ecological investigations. I followed a recursive strategy to develop an image hierarchy using a corrected version of the normalized difference vegetation index (NDVIc) derived from a Landsat ETM+ satellite image over a complex, forested landscape at Lava Cast Forest (LCF), Oregon. At each scale level, I calculated an objective function based on within-object variance and spatial autocorrelation to distinguish between alternative image objects created with the region-merging segmentation algorithm available in the Definiens Developer 7 software. Segmentation quality was considered highest for results exhibiting the lowest overall within-object variance and between-object spatial autocorrelation. I then applied geographical variance analysis to calculate the independent contribution and relative variability of each level in the hierarchy to evaluate the scene's spatial structure across scales. My results reveal overall trends in image object spatial variance consistent with scaling theory, but suggest judging image object quality without sampling the entire range of segmentation parameters is insufficient. Statistical limitations of the spatial autocorrelation coefficient at small sample sizes constrained the number of possible hierarchy levels within the image spatial extent, preventing identification of larger-scale landscape patterns. Geographical variance analysis results show patterns in vegetation conditions at LCF possess a multi-scaled structure. Three levels exhibiting high variance relative to the entire hierarchy coincide with abrupt transitions in the slopes of within-object variance and spatial autocorrelation trends, which I interpreted as scale thresholds potentially important for relating landscape patterns and processes. These methods provide an objective, object-oriented approach for addressing scale issues within heterogeneous landscapes using remote sensing.
Ducey, Craig David, "Hierarchical Image Analysis and Characterization of Scaling Effects in Remote Sensing" (2010). Dissertations and Theses. Paper 399.