Segmentation of remotely sensed images by MDL-principled polygon map grammar

Polygon map grammar -- a generic structure model of polygon maps -- has been developed to an operational level for landuse mapping from remotely sensed images. This grammar enables us to simulate stochastic structures of polygon maps, and thus to simulate ideal and real images of landuse fields. These images provide fully controlled cases for testing image segmentation algorithms. We have developed a general segmentation algorithm which cascades an information-preserving smoothing filter, an edge-preserving smoothing filter, and explicit MDL-based region merging. With both simulated and real images, this algorithm proved to be objective and robust, yielding good results. Using crack edges a mechanism exists and is developed to vectorize raster segmented image to a polygon map data structure -- polyplex. With this data structure, the high-level structure model -- polygon map grammar -- is used to predict missing edges that are not caused by image intensity but are truly not detectable by remote sensors. This paper describes a complete case of this grammar and its application, and a number of basic general mechanisms.