Near infrared spectroscopy is a new on-line analysis technology, which is fit for the detection of organic substance, such as volatile matter of coal． In this paper, 139 coal samples were chosen to acquire diffuse reflectance near-infrared spectra． Then the spectra were correlated to volatile matter of coal with partial least squares． At the same time, different spectral band selecting and pre-treatment methods were used to optimize regression model． At last, the regression model was built for fat coal, and the model result was significantly improved． It shows that regression model of homogeneous set can improve the prediction ability．