To address the class confusion caused by using only spectral features in image clustering for lithological recognition and classification based on multispectral- hyperspectral data, spectral-spatial features are used in multispectral image clustering. Two ways are adopted to include spatial information in clustering. First, multi-scale images are generated using discrete wavelet transform, and are successively clustered from coarse to fine scale images. Second, the spatial class membership probability is calculated using Markov random field model in different image scales, and the combined spectral-spatial class membership probability is then obtained. Finally, the final clustering results are obtained by combining the spectral-spatial clustering results of multiple scales. ASTER multispectral data were used for lithological mapping, and compared with the spectral clustering method based on Gaussian mixture model clustering (spectral clustering), the filtering after spectral clustering, and the spectral-spatial clustering on original image. The results show that the method of spectral-spatial clustering on multiscale image adopted obtained higher accuracies of lithological mapping than the above three comparative methods. Using spectral and spatial information at multi-scales in clustering is an effective method for lithological mapping. The method is applicable to lithological mapping in areas where training samples are difficult to obtain.