Determination of arabica and robusta relative proportion of roasted coffee blends is an important aim for various reasons: different market price, particular organoleptic characteristics, etc. Chemical analysis based on specific compounds is long, costly and not very reliable to determine the proportion of arabica. A spectral database of more than 500 samples (pure arabica, robusta and laboratory blends) has been constituted in diffuse reflectance using a FOSS Nirsystem 6500. The samples chosen for this study came from several arabica and robusta coffee varieties covering different geographical origins. The post-harvest treatments and roasting parameters were representative of production conditions. This base was representative of the spectral variability encountered in commercial coffees. A mathematical model using multivariate regression methods (PLS) was developed to predict the arabica content of blends directly from the NIR spectral data. The model was validated with a set of samples (50) fully independent from the calibration base. The standard error of prediction (SEP) was 2.40% (W/W), the coefficient of determination (R2) of the regression was 0.99 and the regression slope was 1.01. This calibration is robust concerning variability issued from varieties, geographical origins and post harvest treatment. We tested through a specific experimental design crossing 5 levels of blends and 3 levels of roasting (clear, normal and dark) with two replicate by ANOVA the influence of roasting level on the SEP estimation. The accuracy of roasting level was controlled by prediction of luminance value with a NIR calibration developed in our laboratory. The results confirm the ability of the calibration to predict a classical commercial blend. SEP values increase from normal roasting to dark roasting from 1.95% (normal) to 6.4% (dark). There is a significant difference (p < 5%) for the mean predicted value of dark roastings, but no difference exist between clear and norma