<p>The development of <i>spatial knowledge graphs</i> (SKGs) provides a strong data foundation for geospatial applications, such as geographic question answering. However, the incompleteness of SKG hinders the application’s ability to comprehensively understand and reason about relationships between geographic entities. Therefore, predicting links between geographic entities is crucial for advancing geospatial applications. Existing <i>spatial link prediction</i> methods confine the spatial modality of entities to coordinates, thus losing the intrinsic topological information of geographic entities, such as the boundary of polygons. In this paper, we propose a topological relation-aware (TRA) link prediction method for spatial knowledge graphs. We design a topological relation capture module to capture the topological relations of geographic entities. Our method is capable of learning the multi-dimensional semantic information of geographic entities, including points, lines, and polygons, as well as their topological relations. Additionally, we introduce an effective approach to integrate multi-dimensional spatial semantic information with SKG structural information. Extensive experiments on three datasets demonstrate that TRA can learn multi-dimensional spatial features and achieve better performance. Our code and newly constructed geospatial topology (GST) dataset are available at: <a href="https://github.com/NEU-IDKE/TRA">https://github.com/NEU-IDKE/TRA</a></p>