SIZE-CONSTRAINED REGION MERGING: A NEW TOOL TO DERIVE BASICLANDCOVER UNITS FROM REMOTE SENSING IMAGERY

摘要

Landcover maps typically represent the territory as arnmosaic of contiguous units –polygons- that are assumedrnto correspond to geographic entities –like e.g. lakes,rnforests or villages-. They may also be viewed asrnrepresenting a particular level of a landscape hierarchyrnwhere each polygon is a holon –an object made ofrnsubobjects and part of a superobject. The focal levelrnportrayed in the map is distinguished from other levelsrnby the average size of objects compounding it.rnMoreover, the focal level is bounded by the minimumrnsize that objects of this level are supposed to have.rnBased on this framework, we have developed arnsegmentation method that defines a partition on arnmultiband image such that I) the mean size of segmentsrnis close to the one specified; ii) each segment exceedsrnthe required minimum size; and iii) the internalrnhomogeneity of segments is maximal given the sizernconstraints. This paper briefly describes the method,rnfocusing on its region merging stage. The mostrndistinctive feature of the latter is that while the mergingrnsequence is ordered by increasing dissimilarity as inrnconventional methods, there is no need to define arnthreshold on the dissimilarity measure between adjacentrnsegments.