Landscape similarity, retrieval, and machine mapping of physiographic units

• We propose a method for quantitative assessment of similarity between landscapes.
• Landscape search engine finds all similar landscapes in the study area.
• Automatic delineation of physiographic units
• Physiographic provinces in Poland are delineated and compared to reference map.
• GeoWeb application of landscape search over Poland is available online.

We introduce landscape similarity — a numerical measure that assesses affinity between two landscapes on the basis of similarity between the patterns of their constituent landform elements. Such a similarity function provides core technology for a landscape search engine — an algorithm that parses the topography of a study area and finds all places with landscapes broadly similar to a landscape template. A landscape search can yield answers to a query in real time, enabling a highly effective means to explore large topographic datasets. In turn, a landscape search facilitates auto-mapping of physiographic units within a study area. The country of Poland serves as a test bed for these novel concepts. The topography of Poland is given by a 30 m resolution DEM. The geomorphons method is applied to this DEM to classify the topography into ten common types of landform elements. A local landscape is represented by a square tile cut out of a map of landform elements. A histogram of cell-pair features is used to succinctly encode the composition and texture of a pattern within a local landscape. The affinity between two local landscapes is assessed using the Wave-Hedges similarity function applied to the two corresponding histograms. For a landscape search the study area is organized into a lattice of local landscapes. During the search the algorithm calculates the similarity between each local landscape and a given query. Our landscape search for Poland is implemented as a GeoWeb application called TerraEx-Pl and is available at http://sil.uc.edu/. Given a sample, or a number of samples, from a target physiographic unit the landscape search delineates this unit using the principles of supervised machine learning. Repeating this procedure for all units yields a complete physiographic map. The application of this methodology to topographic data of Poland results in the delineation of nine physiographic units. The resultant map bears a close resemblance to a conventional physiographic map of Poland; differences can be attributed to geological and paleogeographical input used in drawing the conventional map but not utilized by the mapping algorithm.