Using nonlinear partial differential equations (PDE) is a recent trend in image processing. In this paper, we propose a new method for extracting spatio-spectral information from hyperspectral satellite (HS) images for classification tasks, which is based on nonlinear diffusion PDEs. As the first step of the method, a feature reduction algorithm is applied to HS data to prevent the final feature vector from being too long. Kernel based principal component analysis is shown to be a good choice for the feature reduction step. At the next step of the method, a set of PDE smoothing filters are applied to a few KPCA components of the HS data, obtained from the first step, which produce a multiscale representation of the HS data. The members of the filter set only differ in their smoothing characteristics which are controlled by two distinct parameters. The extracted features (called extended PDE profile) are then fed into an SVM classifier. Some experiments are conducted on a well-known HS image, Indian Pines data set. Moreover, the classification results of the proposed method are compared to those of some recent spatio-spectral HS image classification methods. The experimental results show the good capability of the proposed method for HS image classification, as well as its better performance than the competing methods.