Probabilistic morphological modeling of hydrographic networks from satellite imagery using Self-Organizing Maps

Adequate and concise representation of the shape of irregular objects from satellite imagery is a challenging problem in remote sensing. The conventional methods for cartographic shape representation are usually inaccurate and will provide only a rough shape description if the description process is to be fully automated. The method for automatic cartographic description of water basins presented in this paper is based on Self-Organizing Maps (SOM) - a class of neural networks with unsupervised learning. So-called structured SOM with local shape attributes such as scale and local connections of vertices are proposed for the description of object shape. The location of each vertex of piecewise linear generating curves that represent skeletons of the objects corresponds to the position of a particular SOM unit. The proposed method makes it possible to extract the object skeletons and to reconstruct the planar shapes of sparse objects based on the topological constraints of generating lines and the estimation of local scale. A context-dependent vertex connectivity test is proposed to enhance the skeletonization process. The test is based on the Markov random chain model of vertices belonging to the same generating line and the Bayesian decision-making principle. The experimental test results using Landsat-7 images demonstrate the accuracy of the proposed approach and its potential for fully automated mapping of hydrological objects.