This work proposes a novel vision-based approach for classifying terrain types around a planetary rover in order to improve autonomous mobility in unknown planetary environment. The key technique of the classification scheme is an identification algorithm for a spatio-temporal texture appearing in motion image sequence, called a Dynamic Texture. It is applied to video sequences which are acquired from the rover's on-board camera. In order to decrease the computational load for estimating Dynamic Texture models with the original method called PCAID algorithm, this work proposes a method combining a system identification method called N4SID with an image compression technique called 2-Dimensional Cosine Transform (2-D DCT). Different types of terrain image sequences are recognized with measurement metrics for the estimated dynamical models. In this paper, some representative metrics are applied to synthetic image sequences, and it is discussed which one is appropriate to distinguish different terrain textures as well as the vehicle's motion paramters such as a taranslational velocity.