Due to the harsh environments, polar region is now becoming one of the typical scenarios for mobile robots. As one kind of representative polar robots, the long-range polar rover robots require higher autonomy to perceive surrounding environments. To improve the real-time performance and environment model accuracy of environment modeling of outdoor environments, multi-scale 2.5D probabilistic elevation grids is proposed. From the perspective of space scale, the model lattices the environment into refined and coarse grids to adapt to different model accuracy requirements and real-time requirements. From the perspective of probability, the refined grids are estimated by Kalman filter while the coarse grids are represented by the statistics of the covered refined grids. Meanwhile, interpolation of the environment model based on hypothesis tests with the presence of ice cracks is studied. The experiments validate the effectivity of the proposed environment modelling and interpolation algorithms.