Rendering of terrain meshes has been an active field of research for several years. Real-time, out-of-core terrain rendering systems based on a hierarchy of level of detail representations, in which each node contains a simplified version of the terrain geometry of its sub-nodes, are available today. Although providing good results at close view distances, these techniques are not well suited to the rendering of distant views of urban models. Once the error-threshold exceeds the building size, they are removed by the simplifier and only represented in the terrain texture. This results in a loss of viewing-direction dependent information and thus a change of the appearance of the city. To overcome this problem, we propose the use of surface light fields as a view-dependent, image based level of detail representation for far distances. We compare two compression techniques, the Per Cluster Factorization and the Linear Mode-3 Tensor Approximation, and we show that the surface light fields are more compact, can be rendered at higher frame rates, and show less aliasing artifacts than the geometry they represent.