Neural Radiance Fields with Physically Based Reflectance for Satellite Images

Abstract. Characterising planetary surface optical properties from satellite imagery is key to understanding surface composition and texture. Neural Radiance Fields (NeRF) have recently demonstrated strong potential for 3D scene reconstruction and novel view synthesis, yet most existing formulations rely on empirical or Lambertian reflectance assumptions that limit their physical interpretability. In this work, we extend BRDF-NeRF by replacing its empirical RPV reflectance model with the Hapke model, a physically grounded BRDF widely used in the planetary remote sensing community to characterise surface scattering properties. Our architecture integrates the Hapke rendering equation into a NeRF backbone and jointly estimates four physically interpretable surface parameters - single-scattering albedo w, phase function parameters b and c, and photometric roughness θ - directly from a sparse set of satellite images. We evaluate the capacity of the framework to recover these parameters against two sources of ground truth: Hapke parameters estimated by a competitive inversion method, and reflectance spectra measured in laboratory conditions using a spectroradiometer. Our experiments show a high correlation between the estimated and reference single-scattering albedo, while the roughness parameter θ is systematically underestimated. In terms of 3D reconstruction and novel view synthesis, our Hapke-NeRF performs comparably to BRDF-NeRF with the RPV model, demonstrating that physically grounded reflectance can be incorporated into NeRF without sacrificing geometric accuracy.