Some components as well as starch properties influence cooking and eating quality of rice. Routine methods to analyse these characteristics are developed on ground milled rice and are time consuming. A large quantity of raw rice (minimum = 50g to 100g) must be milled and then ground before being used for analysis. The present study focused on developing NIRS calibrations to predict amylose and protein content, and gelatinization temperature (GT) from samples at three stages of processing : brown rice, milled rice and ground milled rice. For this purpose, a large and diverse population of rice varieties used by breeders throughout the world and obtained in controlled conditions was used. A total of more than 700 samples were analysed using laboratory reference methods. The NIR spectra of ground milled samples (700) were collected in diffuse reflectance using a FOSS Nirsystem 6500. Additionally, spectra of 506 of the whole milled rices and 255 of brown rices were scanned by NIRS before further processing stages. Partial Least Squares models were used to establish quantitative relations between each NIR database and protein and amylose content and GT values. The performance of calibrations developed for ground milled rice were SECV = 0.15%, r2 = 1.0, RPD (SD/SECV) = 9.5 for Protein content and SECV = 0.99%, r-2 = 0.93, RPD = 2.9 for amylose content. NIRS can also predict protein content from whole milled rices and brown rices spectra with respectively SECV = 0.27% and 0.25%, r2 = 0.99 and 0.98, RPD = 5.5 and 3.8. Calibration models developed for the three stages of rice processing allowed to use them to predict the gelatinization temperature with a good accuracy SECV < 2 and RPD = 3.5 and 2.5 for ground milled rice to brown rice respectively. For amylose content, the performance of calibrations for whole milled and brown rices are lower and can be used only for a screening of varieties for breeders for example. It is concluded that NIRS is suitable for prediction of