Human ability to explore planets (e.g. the moon, Mars) depends on the autonomous mobile performance of planetary exploration robots, so studying on terrain classification is important for it. Vibration-based terrain classification unlike vision classification affected by lighting variations, easily cheated by covering of surface, it analyses the vibration signals from wheel-terrain interaction to classify. Three accelerometers in x, y, z direction and a microphone in z direction were mounted to arm of the left-front wheel. The robot drove on the sand, gravel, grass, clay and asphalt at six speeds, three groups of acceleration signal and one group of sound pressure signal were received. The original signals were dealt using Time Amplitude Domain Analysis. Original data were divided into segments, each segment was a three centimeters distance of driving; eleven features from every segment were normalized. The data from four sensors were merged into a forty-four dimensions feature vector. Ten one against one classifiers of Support Vector Machine (SVM) were used to classify; one against one SVM program from LibSVM was applied to multi-class classification using voting strategy in MATLAB. Facing to the same number of votes, we propose a new algorithm. Experimental results demonstrate the effectiveness of the feature extraction method and the multi-class SVM algorithm.