3D Gaussian Splatting for Detailed Reconstruction of Planetary Surfaces from Orbiter Images

Three-dimensional (3D) mapping of planetary surfaces is critical for exploration missions and scientific research (Gwinner et al., 2016). Previous research mainly focused on employing rigorous techniques such as photogrammetry and photoclinometry to generate topographic products such as digital elevation models (DEMs). While the integration of these two techniques can yield detailed and precise topographic data, photoclinometric algorithms are heavily dependent on radiometric data and surface reflectance behaviors (Chen, Hu, et al., 2024; Liu and Wu, 2023), which limits their use in different circumstances. This paper undertakes a new endeavor to explore the potential of the emerging 3D Gaussian splatting techniques for a detailed reconstruction of planetary surfaces from orbiter images.

Gaussian Splatting has demonstrated outstanding performance in 3D applications for close-range scenes and has recently attracted significant attention. The primary challenge in utilizing 3D Gaussian Splatting for the reconstruction of planetary surfaces from orbiter images lies in the complexity of the planetary push-broom camera models. The sophisticated camera model and projection algorithm complicate this optimization approach. To address this, a two-step approach is proposed to transform the planetary push-broom images into frame-like images. First, photogrammetry is applied to push-broom images to extract precise 3D topography, which is then textured using the corresponding textures from the orthoimages. From the textured 3D landscape, frame images are rendered with careful consideration of overlapping and lighting conditions to better support 3D reconstruction tasks. For surface reconstruction, the 2D Gaussian splatting method (Chen., Li., et al., 2024) is selected and implemented in a coarse-to-fine manner, incorporating a smoothness loss to ensure its suitability for textureless planetary surfaces. In addition to utilizing information from the images, the algorithm also takes into account the camera geometry derived from the previous two steps for improved 3D surface reconstruction.

Experiment analysis is conducted using HiRISE images covering the Jezero crater on Mars. The photogrammetric DEM is generated at a resolution of 1 meter per pixel, and the original images are rectified and mosaicked at their native resolution of 0.25 meters per pixel. A total of 421 frame images are rendered, ensuring high overlapping (e.g., one point appears in eight rendered images) coverages. Compared to the photogrammetric DEM, the DEM generated by 3D Gaussian splatting reveals more subtle topographic details and maintains geometric accuracy.

 

Reference

Chen, D., Li, H., Ye, W., Wang, Y., et al., 2024. PGSR: Planar-based Gaussian Splatting for Efficient and High-Fidelity Surface Reconstruction. arXiv preprint arXiv:2406.06521.

Chen, H., Hu, X., Willner, K., Ye, Z., et al., 2024. Neural implicit shape modeling for small planetary bodies from multi-view images using a mask-based classification sampling strategy. ISPRS Journal of Photogrammetry and Remote Sensing 212, pp. 122-145.

Liu, W.C., Wu, B., 2023. Atmosphere-aware photoclinometry for pixel-wise 3D topographic mapping of Mars. ISPRS Journal of Photogrammetry and Remote Sensing 204, pp. 237-256.

Gwinner, K., Jaumann, R., Hauber, E., Hoffmann, et al., 2016. The High Resolution Stereo Camera (HRSC) of Mars Express and its approach to science analysis and mapping for Mars and its satellites. Planetary and Space Science 126, pp. 93-138.