Exploring the Potential of Refractive NeRFs for Photogrammetric Bathymetry-First Application to UAV-based Data from the Pielach River

Abstract

Accurate 3D reconstruction of underwater environments from above-water photos remains challenging due to refractive distortion at air-water interfaces. This contribution presents the first application of NeRFrac (Zhan et al., 2023) to UAV-based imagery captured in a real-world river area. NeRFrac is a refraction-aware Neural Radiance Field (NeRF) framework that explicitly models the change in direction of light at water surfaces according to Snell’s Law. To adapt NeRFrac to complex outdoor scenes, we introduce a maskbased ray selection that selectively applies refractive modeling only to water-covered regions. We systematically evaluate different indices of refraction and compare global versus local training strategies. The results show that masking improves reconstruction quality in submerged areas, with a physically plausible index of refraction (IOR) of 1.333 yielding the best performance in terms of peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM). While visual differences between masked and unmasked models remain minor, quantitative metrics confirm the effectiveness of the suggested refraction modeling.