Abstract: Levels of detail (LOD) is the key feature of 3D GIS. Aiming at the critical issues of LOD models organization, this paper presents an extended dynamic 3D R-tree structure concerned with LODs, which improves existing node-choosing and node-splitting sub-algorithms. In the node-choosing process, a globally-optimising approach is adopted which is first bottom-up then top-down, and that an improved clustering algorithm based on k-medoids is applied to the node-splitting process. This method makes more even node size, more regular node shape, and less node overlap. Furthermore, a method of integrating R-tree and LOD is put forward based on such kind of good R-tree structure. In order to verify the validity of this method , the simulated experimental analysis results approve that this approach improves 3D query performance greatly relative to existing ones and succeeds in integrating LOD models, which denotes good application prospect and practical value.