The infrared reflectance spectroscopy from the sample simulating natural-rock prepared by kaolin, muscovite and montmorillonite mixed-powders was obtained by a spectrometer. Spectral data preprocessing was done using SNV. Random forest mathematical modeling was used for predicting the components of rock samples. The smallest root mean square error of the predicted three types of rock composition were 0.088 0, 0.095 6 and 0.121 2 respectively. The predictive studies showed that the application of near infrared diffuse reflectance spectroscopy to determining the content of the natural rocks and minerals of various rock composition is feasible. The study provides a theoretical basis for the rapid detection of the rock composition in the future.