A METHOD FOR UNIVERSAL SUPERCELLS-BASED REGIONALIZATION (PRELIMINARY RESULTS)
Keywords: Spatial Patterns, Knowledge Discovery, Regionalization, Supercells, Over-segmentation
Abstract. Geospatial data comes in various forms, including multi and hyperspectral images but also rasters of local composition, local time series, local patterns, etc. Thus, we generalize the SLIC algorithm to work with a library of different data distance measures that are pertinent to geospatial rasters. This contribution includes a description of the generalized SLIC algorithm and a demonstration of its application to the regionalization of the raster of local compositions (of land cover classes). Two workflows were tested, both starting with SLIC preprocessing. In the first, superpixels are subject to regionalization using the graph-partitioning algorithm. In the second, superpixels are first clustered using the K-means algorithm, followed by regions delineation using the connected components labeling. These two workflows are compared visually and quantitatively. Based on these comparisons, coupling of superpixels with a graph-partitioning algorithm is the preferred choice. Finally, we propose using the SLIC superpixel preprocessing algorithm for the task of regionalization of various geospatial data in the same way as it is used for the task of image segmentation in computer vision.