Abstract: This paper introduces an improved method for extracting image texture characters from the large-scale remote sensing image on the basis of retrieval based on image content,which has over 500 000 pixels.This method divides image into sub-images and extracts their texture characters with the method of co-occurrence matrices firstly,then calculates some statistics of these texture characters using a statistical method like Hu's invariable moment.So we can extract a group of texture characters connected with the texture's position in image.It improves the method of co-occurrence matrices and makes the texture characters contain more information.This method can be used in large-scale remote sensing image retrieval and classification.