Dynamic Level-of-Detail 3D Mesh Subdivision for Ellipsoidal Terrains

Sammanfattning

Interactive visualization of planetary terrain requires rendering large geographic areas across a wide range of viewing distances, from global overviews to near-surface perspectives. Since rendering all terrain at maximum resolution is not practical in real time, terrain rendering systems rely on level-of-detail (LOD) methods that adapt geometric detail to the current view. This thesis presents a CPU-driven, cache-oriented LOD method for rendering terrain on an ellipsoidal planetary surface. The method represents the planet using a hierarchical ellipsoidal cube-map structure, where fixed-resolution terrain tiles are selected according to the current camera position and reused through a mesh tile cache. To maintain geometric continuity between neighboring tiles of different resolutions, the method applies localized border stitching during mesh generation. The approach is implemented and evaluated in Carmenta Engine, a geospatial visualization platform. The evaluation examines terrain update cost, cache reuse, screen-space size coverage and stitching overhead under two camera movement patterns and three focal region configurations. The results show that CPU-side tile selection contributes only a small part of the total update cost, while mesh generation and cache reuse have a much stronger influence on runtime performance. Larger high-detail regions improve screen-space size coverage but increase update cost, whereas smaller regions provide lower update times at the expense of reduced detail coverage. The stitching method preserves continuity across the inspected LOD transitions with limited additional cost. The findings indicate that a CPU-driven, cache-oriented tile hierarchy can support real-time planetary terrain rendering under consistent camera movement and high cache reuse. The method is particularly effective when terrain updates can reuse previously generated mesh regions and the localized stitching approach maintains continuity without adding substantial overhead. Although rapid changes in the desired tile configuration can still produce update spikes, the results demonstrate that the proposed architecture is a practical foundation for further development of CPU-managed planetary terrain LOD systems.