The selection of features, including spectral, texture, shape, size, and signal strength, is an important step in computerized information analysis of remotely sensed images. A feature space, which can be generally understood as a multidimensional space consisting of multiple individual features, can be modelled by estimating the distribution of the whole space with prior assumed probability distribution functions (PDFs) once only. However, due to the inter-overlapping phenomenon among points or the confusing influence from surrounding discrete points, it is very difficult to obtain the subtle and procedural structure of the mixture distributions of feature space, and so as to degrade the accuracy and interpretability of the results in further analysis. Extending on the method of Gaussian mixture modelling and decomposition (GMDD), a new feature mining method--stepwise optimization model (SOM) with genetic algorithms (GA) was proposed in this study for the extraction of tree-like hierarchical structure of unknown feature distributions in a feature space. To approximate reality accurately, integration of SOM-GA with symbolic geographical knowledge is essential in the feature mining and classification of remotely sensed images. Knowledge-integrated SOM-GA model that combines the power of SOM-GA and logic reasoning of rule-based inference was therefore proposed. The paper presents conceptual and technical discussions of the model in detail, along with the result of practical application test on a district in Hong Kong region.
No Reference information available - sign in for access.
No Citation information available - sign in for access.
No Supplementary Data.
No Article Media
Document Type: Research Article
State Key Laboratory of Resources and Environmental Information System, Institute of Geographical Sciences and Natural Resources Research Chinese Academy of Sciences Beijing 100101 PR China, Email: [email protected]
Department of Geography the Chinese University of Hong Kong Shatin, N. T. Hong Kong
Publication date: December 1, 2003
More about this publication?