Abstarct: Hyperspectral sensors have more spectral bands than multispectral types, and thus are capable of recording more accurate spectral information of ground surface. This advantage can disturb process of hyperspectral data classification and reduces its accuracy. In addition, Hyperion data as a kind of the hyperspectral images also suffers from noise. In this research, we try to minimize the effects of these problems by means of data mining techniques. Because of the high efficiency of kernel-baxsed methods for high dimensional data classification, the Support Vector Machine (SVM) method was used to classify Hyperion hyperspectral Data. The noise in the Hyperion data was also covered by the implementation of the Robust-SVM – the noise resistant version of SVM. Another way to overcome these problems is to use data reduction techniques. The above mentioned methods and Hyperion image were used to identify alterations of Sarcheshmeh and Darrehzar mines and Sereidun district. In the first step, the use of a feature selection method led to the identification of approximately 11% (18 bands) of the entire spectral bands of Hyperion (165 bands). The results revealed that the Robust-SVM produces more accurate alteration maps than standard SVM. But the most accurate results were obtained when the classification was performed using Robust-SVM and selected bands introduced by the feature selection method. In this case, the Kappa coefficient of classification was 0.61, which compared with the standard SVM using selected spectral bands (0.53), the Robust-SVM with total spectral bands (0.49) and the standard SVM with total spectral bands (0.41) is the most accurate. Also, Support Vector Regression (SVR) was used to perform subpixel classification. It was resulted that the sub-pixel map obtained by the SVR even with the entire spectral bands (i.e. not with the selected spectral bands) and in the standard mode (i.e. not robust SVR) is more precise than the pixel baxsed maps of SVM and Robust-SVM.