Abstract: Based on gray level co-occurrence matrix(GLCM), different multi-scale texture images were extracted, and then the texture bands were combined with original bands to get new composited multispectral bands. The maximum likelihood supervised classification method was applied to new different composited images to test the differences of classification accuracy. Taking Ximen Island as a case study, through analyzing the differences of classification accuracy of different composited images, the following preliminary conclusions were got: (1)the total classification accuracy can be improved from 75.41% to 89.41% by using multi-scale texture images; (2)among different texture indices, correlation and homogeneity texture image calculated from the first band and second band of SPOT-5 image, mean texture image calculated from the third band, homogeneity texture image calculated from the fourth band can both improve classification accuracy than other texture indices images calculated from other bands.

Key words: remote sensing, multi-scale, GLCM, texture