Segmentation-Based Unsupervised Terrain Classification for Generation of Physiographic Maps

Developing an effective method for automatic mapping of physiography is of great interest because such maps have wide range of applications, but creating them manually is expensive and suffers from lack of standards. Many automapping methods have been proposed, but most yield pixel-based maps that do not quite match an appearance and usability of manually drawn maps. In this letter, we propose a method for autocreation of a physiographic map that has handmadelike appearance and functionality. The new method relies on the concept of stacked classification. First, the outcome of existing pixel-based classification is used to construct new features that contain contextual information around each pixel. Second, these new features are used by a segmentation/classification algorithm to create a final map showing generalized landform classes. We describe the design of our method and demonstrate its utility by mapping the physiography of Tharsis region on Mars. A framework of the new method is general enough to improve upon maps created by all previous pixel-based methods. Potential applications include the following: facilitating efficient geologic mapping, enabling computational comparative geomorphology, more effective visualization of topography, and fusion with other data layers within the Geographic Information System framework. The method can also be applied without modification to create segmentation-based maps of land covers.