Enhancing 3D Gaussian splatting with semantic and geometric priors: bridging neural rendering and 3D city modeling

3D Gaussian Splatting (3D GS), known for its efficient and explicit radiance field representation, demonstrates considerable potential for modeling complex 3D scenes. However, its geospatial applicability remains limited, especially for areas such as multi-level scene parsing, heterogeneous 3D data extraction and fusion, and task-driven structured representations. As a preliminary step to address these challenges, this paper proposes a novel 3D GS framework that jointly optimizes geometry and semantics. First, a radiance field optimization mechanism that integrates multi-view 2D semantic labels with monocular depth priors is developed. This mechanism generates Gaussian representations with rich geospatial semantic attributes while substantially improving geometric accuracy. For complex urban environments, a geometry–semantics co-optimization module, comprising a semantics-guided adaptive densification strategy and a depth-weighted semantic propagation method, is further introduced. These strategies effectively suppress cross-view semantic noise and optimize memory efficiency. Experimental results demonstrate that, compared to Light Detection and Ranging (LiDAR)-scanned ground truth, the proposed framework achieves a mean Intersection over Union (mIoU) of 81.2% for semantic segmentation and a mean geometric error of 0.093 m, all while preserving high-fidelity rendering quality. Overall, this work provides a practical pathway for integrating 3D GS into GIS ecosystems and establishes the groundwork for advanced geospatial applications.