An improved spectral similarity measure based on kernel mapping for classification of remotely sensed image

Based on the characteristic of multispectral data,a new function called KSSV is designed in modifying the Gaussian kernel mapping by SSV matching technology.With this function,the feature space of multispectral images could be mapped to high dimension space.Then in the high dimension space,the old similarity measure based on Euclidean distance was replaced by SAM method.In this way,the characteristic information in multispectral images can be exploited adequately and used in many remote sensing applications effectively.At last,the method is applied to unsupervised(k-means clustering) and supervised(minimum distance,SVM) classification experiments.The results show that the classification method with KSSV measure can significantly increase the accuracy of distinguishing between different land types and reduce inconsistency in one category.So the improved method can be more effective in the classification of multi-spectral remote sensing images and achieve better accuracy.