Wood density is an important criterion for material classification, as it is directly related to quality of wood for structural use. Several studies have shown promising results for the estimation of wood density by near infrared spectroscopy. However, the optimal conditions for spectral acquisition need to be investigated in order to develop predictive models and to understand how anisotropy and surface roughness affect the statistics of predictive partial least square regression models. The aim of this study was to evaluate how the spectral acquisition technique, wood surface, and the surface quality influence the ability of partial least square–based models to estimate wood density. Near infrared spectra were recorded using an integrating sphere and fiber-optic probe on the tangential, radial, and transverse surfaces machined by circular and band saws in 278 wood specimens of six-year-old Eucalyptus hybrids. The basic density values determined by the conventional method were then correlated with near infrared spectra acquired using an integrating sphere and fiber-optic probe on the wood surfaces by means of partial least square regressions. The most promising models for predicting wood density were generated from near infrared spectra obtained from the transverse surface machined by the bandsaw, via an integrating sphere ([Formula: see text], RMSEP = 23 kg m⁻³ and RPD = 3.0) as well as for the optic fiber ([Formula: see text], RMSEP = 35 kg m⁻³ and RPD = 2.1). Surface quality affected the spectral information and robustness of predictive models with a rougher surface, caused by band sawing, showing better results.