The existing R-tree spatial clustering technology usually randomly designates or calculates the Euclidean distance between spatial data to select the cluster centers, without considering the topic relevance between spatial data, so that the clustering result is influenced by the initial value of k, and the association between spatial data is only based on geographic location. Aiming at this situation, a method of dynamically constructing spatial Topic R-tree (TR-tree) based on k-means++ was proposed. Firstly, in the traditional k-means++ algorithm, k clusters were dynamically determined by the clustering measure function, and Latent Dirichlet Allocation (LDA) model was introduced into the clustering measure function to calculate the topic probability of each spatial data text, as a result, the topic relevance between spatial data was strengthened. Secondly, the cluster center with the highest probability was selected through the topic probabilities. Finally, the TR-tree was constructed, and the spatial data were dynamically allocated during the construction. Experimental results show that with a slight increase of the R-tree construction time, this method has the indexing efficiency and correlation between nodes significantly improved compared to the algorithm of constructing R-tree based only on geographic location clustering.