We have developed a concept of pattern-based analysis of land-surface where a basic unit of analysis is a pattern of landform elements defined over an arbitrary local region referred to as a scene. The DEM is subdivided into a regular grid of scenes reducing its effective dimension by orders of magnitude. Scenes are described by histograms of local pattern primitive features and similarities between scenes are calculated using histogram distance measures. With scene histograms as cell attribute and histogram distance as a metric the grid-of-scenes can be analyzed much like the ordinary DEM including its segmentation and classification. The result is a framework for efficient and robust automatic classification of topography on continental or global scales using extensive DEM archives. The framework also supports spatial search for similar landscapes. The concept of pattern-based analysis is described and an example pertaining to automatic delineation of physiographic units in Poland is presented. In addition, a GeoWeb application for spatial search of landscapes in Poland is also discussed. To facilitate pattern-based analysis we have developed GeoPAT – a toolbox of GRASS GIS modules intended as a platform for experimentation with the pattern-based analysis of DEMs and other spatial datasets.