Abstract: The classification of soil and rock is more difficult than classification of general terrain surfaces. The traditional methods (supervised classification and unsupervised classification) often yield to low accuracies and poor classification effects when applied to soil and rock classification, a new hierarchical classification algorithm based on characteristic multi-band image is proposed. The new algorithm is a top-down, gradually refinement hierarchical classification method which combines with both advantages of supervised classification and unsupervised classification. The new proposed method achieved the high accurate classification of soil and rock by separating rock and soil step by step hierarchically while making use of several characteristic band groups. Experimental results show that the new proposed method has better performance in improving the classification accuracies, the overall accuracy increases 10% and Kappa coefficient improves 0.1. Also, the new method can overcome “same things with different spectrums” phenomenon effectively.