ABSTRACTTOWARD THE DETECTION OF LANDSCAPE FEATURES:CLUSTERING 3D POINTS USING SPATIAL AND THEMATIC CHARACTERISTICSByBoleslo Edward RomeroThe study of Geography generally concerns phenomena at or near the surface of the earth. High resolutions of 3D quantitative and qualitative data can represent such phenomena as objects or fields. The data can be grouped to reveal representations of contiguous regions of spatial and thematic homogeneity. My thesis is concerned with finding groups of 3D points with similar locations, spatial relationships, and thematic values of spectral reflectance. To accomplish this successfully, I synthesized elements of two geographic theories: point aggregation from cartographic generalization and hierarchical geographic ontology. My experimental design used synthetic 3D point data with spectral values. I employed the multi-dimensional Mean Shift clustering technique from the discipline of Computer Vision, and adapted a 3D range image segmentation accuracy assessment technique. I also contributed new techniques for segmentation quality assessment including two area under the curve indices and the development of new segmentation surface plots. Experimental evaluations included comparisons of the Mean Shift results with K-means clustering results, spatial resolution results, noise evaluation results, and the results of an alternative color configuration. I modified the variable sets to address uneven lighting conditions and employed the experimental methods to grouping real-world terrestrial LiDAR scan data. Though my new spatial relationship variable needs improvement, the methods yielded groups of points representing features in the LiDAR data and provided evidence of the potential for grouping richly attributed 3D points that represent geographic features.
Read
