The image texture was extracted from Landsat TM data using rodogram, a geostatistical function, and then added to multispectral classification for lithological discrimination of an arid region, the Altun Mountains in northwest China. The variogram analysis of the image of the study area indicates that the image has two scales of textures: local and regional textures. Therefore, two different window sizes, 17×17 pixels and 61×61 pixels were chosen to extract textural information using rodogram. The results of image classification show that the classification based on spectral data and geostatistical textural information produced much higher overall accuracy than with the spectral classification alone. Moreover, large window size, at which textural information was extracted and then used for image classification, achieved more accurate result than small window size.