A novel terrain classification technique utilizing motion image sequence taken from planetary rover on-board camera is proposed. The proposed method has an advantage in that it can remotely estimate types of surrounding tarrains, and also has possibility to distinguish properties with dynamic interaction between rover body (wheels) and terrain surface. The characteristics of terrain image sequence is recognized based on a linear dynamical system model called Dynamic Texture. The Dynamic Texture is estimated as a set of parameter matrices, which construct a parameter space such as an observability space. In this paper, experimental results are shown to validate the proposed scheme based on real terrain image sequences obtained from a testbed. And the recognition rates for several distance measures computed from the estimated Dynamic Texture models are evaluated.