An Automated Image-Thresholding Technique for Identifying Generalized Mountain Range Polygons
ISBN 978-85-88783-11-9
Authors
1Raposo, P.
1PENN STATE UNIVERSITY Email: pauloj.raposo@gmail.com
Abstract
Certain topographic features are difficult to identify through scale due to their ambiguous extents, with mountains and mountain ranges being notable examples. Mountain summits are well-defined by points, but identifying the footprint of a mountain or mountain range is difficult and has been the subject of several research projects. Footprint polygons would be useful in production cartography, not necessarily for rendering these polygons themselves, but for delimiting an area in which automated labeling algorithms may place a label as they resolve conflict with other labels. For example, a labeling procedure may curve a label along the medial axis of a skeletonized range polygon while being permitted to deviate from this axis within an acceptable area, defined by the polygon, in order to resolve any label conflicts. Further, polygon area or elevation difference within each polygon (i.e., local prominence) can be used to rank the importance of each mountain range when it is necessary to eliminate some as map scale decreases. We present a method to create such polygons that requires only a DEM in which there is both mountainous and relatively flat area present. Out method uses the technique of maximum entropy binary thresholding (“maxent” thresholding) from photographic image processing, and refines the product of that algorithm using geographic distance-based thresholding. The method produces polygons identifying mountain ranges as those areas of a DEM which maximize the information content of the segmented “mountain” and “non-mountain” topography. Strengths of the method include its basis in objective image segmentation optimizing the product’s representation of vertically prominent and non-prominent land, and that the output polygons do not necessarily follow geographic contours. The latter point is important, since reclassifying DEMs by a certain elevation produces ranges that do not conform with human perception of mountains as locally prominent. The method is fully automated using open-source software libraries (GDAL, OpenCV, NumPy) in the Python programming language. We first calculate the slope raster of a DEM. The range of values of this raster is recorded, and the raster is binary-thresholded (i.e., reclassified into 0 and 32767 for a 16-bit DEM) at the first integer value above the minimum. The entropy of this classified image is calculated. This process repeats for every integer value within the elevation interval, keeping note of which threshold value produces the output of highest entropy. The DEM is then thresholded at the optimal value. This typically produces a mottled or fractured-looking binary image in which the highest binary value (i.e., 32767 in a 16-bit DEM) represents areas of relatively greater incline. To generalize this raster zone, the Euclidean distance of every DEM pixel to the nearest cell of highest binary value is calculated. The average area of the lowest binary cell isolated regions (i.e. small contiguous 0-value cell areas) is calculated (Zero_area), the square root of which (sqrt(Zero_area)) is used to threshold the distance raster. This effectively classifies small areas of “non-mountain” interspersed within large “mountain” areas as part of the mountainous portion of the image (e.g., small intermountain valleys or lakes), generally producing raster regions with much less gaps. These regions are polygonized, and any small areas within the polygons are filled, using Zero_area as a threshold. In the presentation, we explain our implementation in detail and display example products from various mountainous locations around the world using ASTER GDEM data.