Least square support vector machine (LSSVM) combined with successive projection algorithm (SPA) method was applied for near-infrared (NIR) quantitative determination of the octane number in fuel petrol. The NIR spectra of 87 fuel petrol samples were scanned for model establishment and optimization. First order derivative Savitzky-Golay smoother (1st-d'SG) was utilized to improve the NIR predictive ability. Its pretreatment effect was compared with the raw data. SPA was applied for the extraction of informative wavelengths. Considering the linear and non-linear training mechanism, LSSVM regression was employed to establish calibration models. The correlation coefficient (R) and root mean square error (RMSE) were used as the model evaluation indices; accordingly the octane number in fuel petrol was quantitatively determined with the prospective predictive indices. Results showed that after pretreated by 1st-d'SG, 8 SPA-selected wavelengths was generated as the inputs of LSSVM, so that the calibration models were optimized in the way of combining the SPA-LSSVM regression with the SG smoother. The prediction results were quite satisfactory, with the calibrating correlation coefficient of 0.951 and the RMSE of 3.282. An independent testing sample set was used to evaluate the optimal model, the testing correlation coefficient was 0.903 and the RMSE was 4.128. We conclude that NIR spectrometry is feasible to determine the octane in fuel petrol by establishing SPA-LSSVM models. The 1st-d'SG pretreatment has the advantage of selecting wavelengths containing the implicit information. The combination of 1st-d'SG pretreatment and SPA-LSSVM show its applicable potential to predict the octane number in fuel petrol.