Traditional approaches to generate digital elevation model（DEM) from aerial imagery consist of two steps. The first step establishes feature correspondences and determines their height

and the second interpolates height to generate dense DEM. Because the first step does not apply global optimization and the second step usually introduces interpolation error

they impair the quality of DEM. This paper describes DEM as Markov random fields

formulates DEM generation as pixel labeling. It generates DEM in a global optimization framework and does not need interpolation. Then

this paper constructs multi-resolution Height Fields and proposes a multi-level pixel labeling strategy. It determines the Height Fields on the highest level at first

and then determines the Height Fields on the rest levels step by step. It improves efficiency greatly. At last

this paper modified Belief Propagation algorithm to determine Height Fields on a specific level. It passes Height Fields on the higher level to the lower level

restricts the possible height and reduces the search space greatly. As a result

it improves both efficiency and quality. Experimental results have shown that high quality DEM have been generated by the proposed approach.