Making a study of the calculating methods of the Within-Class and Between-Class Scatter matrices in the Hyperspectral images Mohsen Ghamary Asl MA Student of Remote Sensing K.N.T. University, Iran Email: m_ghamary@yahoo.com Scatter matrices are used in Feature Extraction (for dimension reduction of the space) methods, that will be very useful and usable for Classification of the Hyperspectral images. To find out the scatter matrices, statistical parameters such as mean and covariance are required which should be estimated using training samples. The first idea is to use methods related to Multispectral images. But these approaches can only be useful if there are great numbers of training samples in a single class compared to the number of space dimensions (This is often the case for Multispectral images). For Hyperspectral images, on the other hand, it does not happen. In other words, in Hyperspectral images, the number of training samples is less than that of dimensions. Therefore, there is a serious need for new ways of determining and estimating the statistical parameters mentioned above. Parametric methods are employed to estimate statistical parameters in Hyperspectral images. The nature of these images is that there is a great amount of correlation between spectral bands, resulting in each class having several clusters. On the other hand, each cluster, in turn, has its own statistical parameters (mean and covariance) and in fact each class would be a combination of different distributions. So it is necessary to point out that we would not be able to fully determine data distribution of a class and calculate the appropriate scatter matrix just using a mean and a covariance value. As a result, we intend to put the statistical parameters away and try to make use of Non-parametric approaches in order to determine scatter matrices since in this way scatter matrices are calculated directly by the available data. The main Non-parametric methods currently in use are NDA and NWFE. Various experiments and assessments along with results obtained show that NWFE is the best method for calculating the within-class and between-class scatter matrices and feature extraction in Hyperspectral images.