The fast and reliable quality determination of alfalfa crop is of interest for producers to make management decisions, the dealers to determine the price, and the dairy producers for livestock management. In this study, the crude protein (CP), one of the main quality indices of alfalfa, was estimated using the visible and near-infrared (Vis/NIR) spectroscopy. A total of 68 samples from various variety trials of alfalfa crop were collected under the irrigated and rainfed conditions. The diffuse reflectance spectral data of the undisturbed alfalfa stem and leaves were collected in the wavelength range of 400-2500 nm using a portable fiber optic spectroradiometer. Different spectral pretreatments were implemented to eliminate the irrelevant information from spectral data. Partial least squares regression (PLSR) method was then used to extract the CP predictive models using the spectral data as independent variables and the CP values obtained from the standard analytical method as dependent one. Primary results showed that the wavelength region of 400-900 nm had the better predictive ability in comparison with the 900-2500 nm spectral region. In the 400-900 nm region, the best model was obtained from multiplicative scatter correction (MSC) preprocessing with a correlation coefficient in leave-one-out cross validation (r_cv) of 0.758 and a root mean square error in leave-one-out cross validation (RMSECV) of 1.096%. While, the best model in the spectral region of 900-2500 nm resulted from 2^nd derivative pretreatment with an r_cv and RMSECV of 0.627 and 1.309%, respectively. When the whole spectral region (400-2500 nm) was used, no improvement in the predictive ability of PLS models was achieved in comparison with 400-900 nm region, where the best model resulted in r_cv and RMSECV values of 0.757 and 1.098%, respectively. Collecting the data during the subsequent years and increasing the harvesting times in each year can improve the robustness of CP predictive models by enhancing the variability in CP value.