Efficient Shallow Water 3D Reconstruction Using UAV Bathymetric LiDAR and Multi-View Imagery via Gaussian Splatting

Duration: 6 months
Completion: December 2025
Supervisor: Dr. Lida Asgharian Pournodrati
Examiner: Prof. Dr.-Ing. Uwe Sörgel

Accurate three-dimensional reconstruction and realistic novel view synthesis from aerial imagery are essential for applications such as coastal monitoring, environmental analysis, navigation safety, and digital twins of natural environments. In recent years, neural rendering approaches such as Neural Radiance Fields (NeRF) [1] and 3D Gaussian Splatting (3DGS) [2] have demonstrated remarkable performance in photorealistic scene reconstruction from multi-view imagery. Despite these advances, reconstructing coastal and shallow-water environments remains particularly challenging due to complex light–water interactions, including reflection, refraction, and wavelength-dependent attenuation.

A major limitation of existing reconstruction pipelines arises during the initialization stage. Classical photogrammetric methods such as Structure from Motion (SfM) [3] and Bundle Adjustment (BA) [4] often fail over water surfaces because of their textureless and dynamic nature. As a result, sparse point clouds generated during initialization are incomplete or entirely missing in water-covered regions. When such sparse reconstructions are used as input for neural rendering methods, the optimization process becomes unstable, leading to inaccurate, fragmented, or incomplete 3D reconstructions. Furthermore, reconstruction methods that rely primarily on photometric consistency are inherently underconstrained in low-texture regions, which further degrades geometric accuracy.

Coastal environments introduce additional challenges due to the optical properties of water. Refraction at the air–water interface causes systematic geometric distortions of submerged structures, while surface reflections and light scattering reduce image quality and consistency across views. These effects violate the assumptions of straight-line light propagation commonly used in NeRF- and 3DGS-based approaches, making accurate underwater reconstruction particularly difficult without explicit modeling.

This thesis addresses these challenges by proposing an enhanced 3D reconstruction framework based on Gaussian Splatting that integrates UAV-based bathymetric LiDAR data with multi-view aerial imagery. The proposed approach focuses on large-scale 3D reconstruction from airborne imagery and improves the robustness of the initialization stage by introducing a novel strategy for adding additional points alongside SfM-derived geometry. By integrating 3D LiDAR data with SfM points, the method significantly improves geometric accuracy in challenging shallow-water and coastal environments. Furthermore, a novel position correction method for 3D Gaussians is introduced to enhance spatial consistency. To stabilize the optimization process and improve reconstruction quality, the loss function is extended with geometry-aware regularization terms, leading to improved densification, completeness, and precision of the final 3D reconstruction. In addition, a refraction-based correction of underwater Gaussian positions is incorporated, compensating for light bending at the air–water interface using Snell’s law. In contrast to most existing neural reconstruction approaches, which primarily target close-range or terrestrial scenes, this work emphasizes large-scale airborne reconstruction of shallow-water and coastal environments.

Multi-view UAV imagery is first processed using Structure from Motion (SfM) to estimate camera poses and generate an initial sparse point cloud. A pretrained U-Net model predicts a water mask to identify regions affected by water-related distortions. To address SfM failures over textureless water surfaces, additional points are introduced through a water surface completion strategy, improving the robustness of the initialization. Bathymetric LiDAR data is then integrated to enhance geometric accuracy in shallow-water regions. During optimization, geometry-aware regularization terms based on depth, surface normals, and scale are applied to stabilize training. For underwater areas, a refraction-aware position correction compensates for light bending at the air–water interface. Together with adaptive density control, the method produces accurate and complete 3D reconstructions from airborne imagery.

The proposed method is evaluated on multiple real-world UAV datasets covering shallow-water and coastal environments. Experiments compare the proposed framework with baseline 3D Gaussian Splatting methods using both qualitative visual inspection and quantitative metrics. The evaluation assesses reconstruction completeness, geometric accuracy, and rendering quality under challenging conditions such as textureless water surfaces and strong reflections.

The proposed method is evaluated on four UAV datasets: Hartheim, Autobahnsee, Kuhsee, and Stavanger. Quantitative results show that the proposed approach consistently outperforms standard Gaussian Splatting and other baselines across all datasets, achieving higher PSNR and SSIM values and lower LPIPS scores. Improvements are most evident in water-dominated scenes, where baseline methods suffer from incomplete initialization and geometric instability.

Qualitative 3D reconstruction comparisons further demonstrate the benefits of the proposed method. The reconstructed scenes are denser and more complete, with fewer holes and reduced artifacts in shallow-water regions. Structural details are better preserved, and the overall geometry appears smoother and more consistent across viewpoints.

Supplementary videos confirm these findings by demonstrating improved novel-view rendering, smoother camera motion, and reduced visual artifacts compared to baseline methods. Overall, the results highlight the effectiveness of integrating LiDAR data, geometry-aware regularization, and refraction-aware correction for accurate large-scale airborne reconstruction of shallow-water and coastal environment.

This thesis presents an enhanced 3D reconstruction framework based on Gaussian Splatting for shallow-water and coastal environments using airborne data. By integrating bathymetric LiDAR with multi-view UAV imagery, improving initialization through additional geometry, introducing geometry-aware regularization, and applying refraction-aware correction of underwater Gaussian positions, the proposed method achieves more accurate and complete 3D reconstructions. Experimental results across multiple real-world datasets demonstrate consistent improvements in geometric completeness, accuracy, and visual fidelity compared to standard Gaussian Splatting. The proposed approach highlights the potential of combining LiDAR sensing with neural Gaussian representations for robust large-scale airborne reconstruction in challenging aquatic environments.

[1] B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng,
“NeRF: Representing scenes as neural radiance fields for view synthesis,” Proceedings of the European Conference on Computer Vision (ECCV), pp. 405–421, 2020.

[2] B. Kerbl, G. Kopanas, T. Leimkühler, and G. Drettakis, “3D Gaussian splatting for real-time radiance field rendering,” ACM Transactions on Graphics (SIGGRAPH), vol. 42, no. 4, ACM, 2023, pp. 1–14. [Online]. Available: https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/

[3] J. L. Schönberger and J.-M. Frahm, “Structure-from-motion revisited,” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4104–4113, 2016.

[4] B. Triggs, P. F. McLauchlan, R. I. Hartley, and A. W. Fitzgibbon, “Bundle adjustment – a modern synthesis,” Vision Algorithms: Theory and Practice, ser. Lecture Notes in Computer Science, B. Triggs, A. Zisserman, and R. Szeliski, Eds., Springer, vol. 1883, pp. 298–372, 2000.