BRLO-Tree (Block-R-Tree-Loose-Octree) is presented in this paper based on the R-Tree and Loose-Octree. The aim of the structure is to visualize the large scale and complex dynamic scenes in a 3D (three-dimensional) GIS (Geographic Information System). A new method of clustering rectangles to construct R-tree based on an improved K-means algorithm is put forward. Landform in 3D GIS is organized by R-Tree. The block is used as the basic rendering unit. The 3D objects of each block are respectively organized by a Loose-Octree. A series of techniques, based on this data structure, such as LOD (Level of Detail), relief impostors are integrated. The results of the tests show that BRLO-Tree cannot only support the large scale 3D GIS scene exhibition with wandering and fighting, but it can also efficiently manage the models in a dynamic scene. At the same time, a set of integrated techniques based on BRLO-Tree can make the rendering pictures more fluence and the rendering time vastly improved.