In the development of positron emission tomography (PET) detectors, understanding and optimizing scintillator light collection is critical for achieving high performance, particularly when the design incorporates depth-of-interaction (DOI) encoding or time-of-flight information. Monte-Carlo simulations play an important role in guiding research in detector designs and popular software such as GATE now include models of light transport in scintillators. Although current simulation toolkits are able to provide accurate models of perfectly polished surfaces, they do not successfully predict light output for other surface finishes, for example those often used in DOI-encoding detectors. The lack of accuracy of those models mainly originates from a simplified description of rough surfaces as an ensemble of micro-facets determined by the distribution of their normal, typically a gaussian distribution. The user can specify the standard deviation of this distribution, but this parameter does not provide a full description of the surface reflectance properties. We propose a different approach based on 3D measurements of the surface using atomic force microscopy. Polished and rough (unpolished) crystals were scanned to compute the surface reflectance properties. The angular distributions of reflectance and reflected rays were computed and stored in look-up tables (LUTs). The LUTs account for the effect of incidence angle and were integrated in a light transport model. Crystals of different sizes were simulated with and without reflector. The simulated maximum light output and the light output as a function of DOI showed very good agreement with experimental characterization of the crystals, indicating that our approach provides an accurate model of polished and rough surfaces and could be used to predict light collection in scintillators. This model is based on a true 3D representation of the surface, makes no assumption about the surface and provides insight on the optical behaviour of rough crystals that can play a critical role in optimizing the design of PET detectors. This approach is also compatible with existing simulation toolkits and next steps include the implementation in GATE.