PRISM-Splat: Physically rendered in-water semantic model for 3D scene reconstruction

Underwater 3D scene reconstruction is highly challenging due to severe light scattering, colour attenuation, and the presence of dynamic elements such as swimming fish. In this paper, we present PRISM-Splat, an efficient and robust 3D Gaussian Splatting framework driven jointly by physical models and semantic features. To fundamentally separate the scattering medium from the underlying scene geometry, we introduce a physics-driven hybrid Gaussian representation derived from the radiative transfer equation (RTE). Rather than applying rigid geometric constraints, we propose a confidence-aware depth prior that adaptively guides the structural optimisation, effectively preventing geometric collapse in severely degraded regions. Furthermore, to address the persistent issue of dynamic underwater distractors, we integrate a multi-view semantic feature field that intrinsically filters out moving objects without requiring explicit priors. Extensive experiments on both authentic and simulated datasets demonstrate that PRISM-Splat achieves superior performance in high-fidelity 3D reconstruction and accurate colour restoration compared to state-of-the-art methods.