Enhancing Multi-Temporal 3D Reconstruction of River Corridor Dynamics with Deterministic NeRF Sampling Strategies

Effective monitoring of hydro-geomorphological processes such as sediment displacement and channel migration using UAV-captured RGB images requires accurate, temporally consistent 3D reconstructions that enable the estimation of height/volume changes across multiple acquisitions epochs, despite variations in imaging conditions, flight paths, and camera geometries. Neural Radiance Fields (NeRF) offer a powerful framework for reconstructing complex fluvial environments from multi-view imagery. However, standard NeRF training relies on equal‑probability random pixel sampling, a strategy that introduces stochasticity into the reconstruction process and undermines the temporal repeatability required for change detection. When separate NeRFs are trained for different epochs the random sampling of pixels – combined with differing camera poses – leads to inconsistent ray distributions and surface sampling, and reduced comparability of resulting reconstructions. These effects are particularly detrimental in dynamic river corridors where low-texture surfaces, water reflections and surface motions already limit reconstruction coverage and stability.

We propose Coverage-Efficient Non-redundant Sampling (CENS) as a grid-based deterministic alternative to random pixel sampling in NeRF training, combined with a pose-consistent inference framework, to improve temporal reconstruction repeatability for hydro-geomorphic change analysis. Instead of randomly sampling pixels per iteration, we impose a spatially deterministic sampling grid that ensures intra-epoch spatial coverage and consistency. To address the challenge of differing camera poses between epochs, we introduce a cross‑epoch evaluation strategy in which all epochs are queried using a shared set of reference camera poses – either drawn from one epoch’s camera set or defined as a synthetic virtual camera lattice – after alignment to a common world coordinate frame. This enables pose‑consistent rendering across time even when the original acquisition geometries differ.

Rendered depth maps from these shared viewpoints are then projected onto a fixed world-space grid, enabling cell-wise comparison of height and volume change. This two-layer determinism—fixed sampling during training and fixed camera/grid sampling during inference—decouples temporal analysis from the stochastic and pose‑dependent variability inherent in standard NeRF pipelines.

We will evaluate the method on multi‑epoch UAV imagery of an active river reach, comparing reconstructions trained with traditional random pixel sampling against those produced using the proposed CENS approach. With improved intra-epoch spatial coverage and geometric stability, combined with pose-consistent inference, we anticipate that the proposed approach will yield more coherent inter-epoch estimates of sediment elevation, channel-bed change and/or bank erosion, with lower uncertainty in low-textured or reflective regions. We expect corresponding 3D points from different epochs to align more consistently in world space, enabling more reliable detection of subtle geomorphic adjustments like bar migration, scour-fill cycles, and/or sediment redistribution.

This study aims to demonstrate that sampling strategy and inference geometry are critical, yet previously overlooked, determinants of temporal repeatability in NeRF-based hydro-geomorphological monitoring. By replacing stochastic pixel sampling with deterministic grids and enforcing shared inference poses across epochs, we introduce a simple but potentially powerful modification that may enhance the reliability of NeRF-derived height/volume change estimates in dynamic river environments. The approach could open new possibilities for long‑term environmental and hydro-geomorphological research where spatiotemporal consistency is essential.