A method of unifying the normal orientation of point clouds in multi-layer Riemannian graphsis presented to address the challenges for existing normal propagation methods of point clouds sampled from curved surfaces in quick processing of massive data. In this method, the point clouds are recursively divided into subsets to obtain the core point sets. The surface variability of the core point sets controls the recurrence number, and a multi-resolution model of tree structure is constructed for the point clouds. The nodes of the point-cloud multi-resolution model are traversed from top to bottom, and the multi-layer Riemannian graph of the point clouds is thus constructed from the subset of non-leaf nodes. Using the sequential traversal method, the normal unification of the sample points in the top-layer Riemannian graph is transmitted downwards. For each Riemannian graph unit, the minimum spanning tree algorithm is used to obtain the normal unification of the sample points. The experimental results demonstrate that this method can effectively improve the computational efficiency and memory utilization in processing massive point clouds and ensure the accuracy of the normal propagation of sample points in complex feature areas.