ИСТИНА

Vector approaches dominate contemporary research in map generalization. For many tasks vector data are more appropriate or convenient, and the approach to their generalization is usually tied to the storage model used. In practice, most research is focused on topographic maps and base maps for web services; raster models are rarely used for these data. Nonetheless, some important cartographic and GIS applications ask for a full-fledged system of raster generalization methods. These come primarily from thematic and special-purpose mapping and geospatial analysis (e.g., statistical surface analysis, climatological and oceanographic visualization and analysis, etc.). For some geographic data the raster model is more natural than vector. Our study aims to take an inventory of past and current research in raster generalization both in GIScience and related fields, reveal problems and uncovered topics, and to offer principles and methods as context to raster generalization systems. We have illustrated that the generalization of rasters is a nuanced procedure that requires attention to several spatial and statistical sources of error. Most notable among these are the effects of projection distortion across large areas, which make supposedly uniform pixels in fact not uniform, and the Modifiable Areal Unit Problem, which confounds statistical aggregation and analysis, such as seen in raster coarsening and resampling. We have begun to address the issues raised by implementing a method using variably-sized kernels which are calibrated using the areal error equations of the map projection used. Further work will address statistical aggregation techniques and methods of measuring structure and information in continuous raster surfaces to enable characterization and evaluation of generalized images.