Reconstruction of riverbed topography with neural radiance fields

This thesis evaluates BathyFacto, a neural radiance field variant that explicitly models light refraction at the water surface, for mapping river beds from drone images of the Pielach River in Austria. Using greedy sequential optimization of preprocessing, extraction and training parameters, the best configuration reached an RMSE of 0.30 m against ground-truth points, compared with 0.02 m for the reference LiDAR. Compared with a refraction-unaware Nerfacto baseline, BathyFacto generally brought the reconstructed riverbed closer to the reference, though overhanging vegetation and increasing water depth remained error sources, and its advantage over the baseline was small on real data.