A new classification and fusion method for multi-source of remote sensing images was put forward based on the D-S evidence theory. Select the interesting region of class by experience and obtain basic probability assignment function by extracting feature at first, then combine multi-source image to be grouped with Dempster's orthogonal rule to get the result of classification. Experiments show that the proposed method is superior to the K-mean. The uncertainty in classification is effectively decreased, as well as the classification accuracy is improved.

Three kinds of clustering methods， including KCN (Kohonen Clustering Network)， FKCN (Fuzzy cMeans based Kohonen Clustering Network) and EPKCN (Evolutionary Programming based Kohonen Clustering Network) that were applied in the classification of remote sensing image， were discussed. Experiments show that these unsupervised learning methods had different characters in classifying land use/cover of remote sensing. To EPKCN， the vision effect of classification is best and the rate of single iteration is fastest; To FKCN， when the training process trends to convergence， the total training rate is fastest. However， taking into count the demand of land use/cover classification in remote sensing， EPKCN is the best one in these three algorithms, and can be applied in unsupervised classification of remote sensing land use/cover.