The original image was properly divided into unoverlapped blocks according to the small change of texture in local areas of the remote senses image, each of which associated with a uniform texture. Then texture feature of each block was formed by calculating the mean and variance of Gabor filtered image. Rotation normalization was realized by circular shift of the feature elements to get the invariant texture feature vector. The classification of image blocks was also completed by using a simple unsupervised clustering algorithm. The experiments of the real images show that the method is effective.