The present paper uses a combination method of wavelength selection and outlier spectra detection for quantitative analysis of nature gas combustion process based on its near infrared spectra. According to the statistical distribution of partial least squares (PLS) model coefficients and prediction errors, the method realized wavelength selection and outlier spectra detection, respectively. In contrast with PLS, PLS after leave-one-out for outlier detection (LOO-PLS), uninformative variable elimination by PLS (UVE-PLS) and UVE-PLS after leave-one-out for outlier detection (LOO-UVE-PLS), the root-mean-squared error of prediction (RMSEP)based on the method for CH4 prediction model is reduced by 14.33%, 14.33%, 10.96% and 12.21%; the RMSEP value for CO prediction model is reduced by 67.26%, 72.58%, 11.32% and 4.52%; the RMSEP value for CO2 prediction model is reduced by 5.95%, 19.7%, 36.71% and 4.04% respectively. Experimental results demonstrate that the method can significantly decrease the number of selected wavelengths, reduce model complexity and effectively detect outlier spectra. The established prediction model of analytes is more accurate as well as robust.