We propose new R-tree construction techniques (CSR-tree) for spatial databases. The main ideal of this algorithm is to make the spatial objects that near to each other in spatial space in nearest leaf nodes, and to reduce the overlap among the spatial objects' rectangles. Given a collection of multi-dimensional spatial objects with rectangles, we cluster them to k groups by distance relativity, sort all the spatial objects in the i-th (iepsi[1,k]) group, and then sort all the groups by the group center points, and build the R-tree bottom-up at last. We proposed and implemented several variations and performed experiments on synthetic 3D data. The experimental results show that the CSR-tree outperforms the previously R-tree methods in query efficiency and space utilization.