A New Method for Feature Mining in Remotely Sensed Images

Extending on the method of regression-class mixture decomposition (RCMD), a RCMD-based feature mining model with genetic algorithm (coined RFMM-GA) is proposed in this paper for the extraction of features in complex remotely sensed images with a large proportion of noise. Within the framework of RFMM-GA, different features in the feature space correspond to different components of a mixture in which each of its components can be specified by a certain type of parametric distribution and the suitable parameter sets. The model captures nicely the overlapping and noisy conditions usually encountered in remotely sensed images. Features are successfully mined when the corresponding parameter sets are appropriately estimated. Through the embedded GA, features with the assumed components are hierarchically mined until the data set is decomposed into a group of feature patterns. Compared to conventional methods, the RFMM-GA has several distinct advantages: (1) The initial number of features does not need to be specified Geoinformatica (2006) 10: 295–312 DOI 10.1007/s10707-006-9829-6 Y. Leung (*) Department of Geography and Resource Management, Center for Environmental Policy and Resource Management, and Joint Laboratory for Geoinformation Science, The Chinese University of Hong Kong, Hong Kong, China e-mail: [email protected] J.-C. Luo :D.-P. Ming LREIS, Institute of Geographical Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing, 100101, China J.-C. Luo e-mail: [email protected] D.-P. Ming e-mail: [email protected] J.-H. Ma Department of Mathematics and Information Science, Chang_an University, Xi_an, 710064, China e-mail: [email protected] a priori. The procedure terminates after all relevant features have been unravelled. (2) Large proportion of noisy data in the mixture can be tolerated. (3) Parameter estimations of individual features are virtually independent of each other. (4) Variabilities in shapes and sizes of the features in the mixture are accounted for. Three experimental results on the extraction of ellipsoidal and linear features demonstrate the effectiveness of the RFMM-GA model for feature mining in noisy data with mixed feature distribution.