Near Infrared (NIR) spectrum was used to detect the fat content in rice. NIR spectra of 90 rice samples were measured. Kennard-Stone method was used to select the calibration set and prediction set samples. The effects of different pretreatment (normalize, first derivative and second derivative methods) have been compared for the accuracy of the models. The best pretreatment method is the first derivative. The competitive adaptive reweighted sampling was applied to screening the key wavelengths associated with the sample properties. Finally, thirty key wavelengths are selected by Multiple Linear Regression further. The most typical key wavelengths are 1343 nm, 1489 nm and 1583 nm which related to the groups of -CH and -OH in rice fat. The detection model of fat content of rice based on near infrared spectroscopy has higher precision whose coefficient of determination, root mean square error of calibration and relative deviation are 0.9589, 0.2236 and 5.53%.