This project aimed to classify multi-spectral satellite images using selective real image bands and principal component images. The problem comes from selecting the type and the number of bands for classification process. The widely used approach for selecting the best bands based on those bands that have the largest sum of squared principal axes that account for the largest total variance derived form the variance covariance matrix for the multi-spectral images of the scene. In the present work, this technique has been examined by selecting the number and the type of bands with the largest diagonal elements in the variance co-variance matrix. The volume of the ellipsoid associated with a particular band triplet is a constant multiple of the square root of the product of the eigenvalues for 3x3 variance co-variance matrix of that triplet The trace of the variance co-variance is invariant under rotation. Under rotational transformation, the product of the eigenvalues is equal to the determinant of the original 3x3 sub-matrix. Therefore, the band triplet that provides the ellipsoid of maximum volume have been investigated by computing and ranking in order the determinants of each 3x3 principal sub matrix of the original matrix. Supervised and un-supervised classification techniques have been used to perform the classification process. Maximum likelihood method is adopted as supervised classification. This method is applied on the original image bands and on the principal component images. Our results show that the classification accuracy increased with increasing the number of selected bands until it reached to the case that the classification accuracy is not affected by increasing the selecting number of the spectral bands. The classification accuracy increased from 72% to 88.7% with increasing the number of bands. The triplet original bands (3. 5, and 6) investigate 86.9% classification accuracy. Our results also show that the overall accuracy with different selecting number of principal component images ranging between (87%- 89%). K-mean method is adopted as unsupervised classification. This method depends on the initial chosen seeds of the cluster configuration. Since, the initial seeds are chosen randomly or the user supplies a set of means, or cluster centers in the n dimensional space. New method is suggested to choose the initial seeds randomly with the maximum distance between the centers of the chosen clusters. New method is also suggested as unsupervised classification method, namely "class method". In this method, the number of classified regions depends mainly on the number of chosen bands for classification purpose. The label of each class is defined by linear combination with number of threshold values (number of chosen bands). The threshold values have 0 or 1 value. The main feature of this method is efficient and fast. It should be mentioned that this study has been concerned with the satellite multi band images, thematic Mapper (TM) images and devoted for classifying image regions over the west of Iraq; namely Ramadi region.