Abstract: In order to develop study of land surface temperature prediction service, and improve forecast accuracy of daily surface temperature, based on the numerical forecast products of ECMWF and T213, daily land surface temperature from 2007 to 2012 in Fushun City, using the methods of stepwise regression and BP neural network model to build the forecast model of land surface temperature in Fushun City and test the accuracy of the model. The results showed that: height field, sea level pressure field, temperature field of ECMWF and divergence, height field, sea level pressure field, pressure field, K index, water vapor flux, relative humidity, temperature, ground temperature field, vorticity field of T213 were a significant correlation with land surface temperature. Accuracy test of the prediction model showed that the surface the average temperature and the surface minimum temperature forecast effect was good, the error of less than 3℃ forecasting accuracy achieves more than 79%. The comparison of 2 models showed that BP neural network prediction model was better than the stepwise regression prediction model and the stability of the stepwise regression prediction model was better than BP neural network prediction model.