Abstract:Based on air temperature observation data of automatic weather stations in Shaanxi Province, by using sliding training period and unitary linear regression methods, air temperature from CMA Land Data Assimilation System (CLDAS)〖JP2〗 is tested and corrected at stations, then the correction coefficient is interpolated〖JP〗 to grid point to revise CLDAS temperature at the grid. Finally, temperature forecast from European Centre for Medium Range Weather Forecasting (ECMWF) is corrected by respectively using the air temperature data from observation and CLDAS before and after correcting, which is tested by non independent and independent validations. The results show that spatial distribution characteristics of temperature from CLDAS are basically consistent with those from station observations, but there are some errors between them, which can be reduced by correcting with observation data. The accuracies of absolute error less than 1℃ or 2℃ of CLDAS temperature before correcting are respectively less than 20% and 30% in Qinling and Daba Mountains, where accuracies are smaller than in other areas and are improved by more than 40% after correction. The CLDAS grid temperature data after correcting are used to correct the temperature prediction of ECMWF improving the accuracy of the model forecasts. It also improves the temperature forecast quality of ECMWF in the areas with few weather stations, and is suitable for high resolution meteorological grid forecast. The accuracies of absolute error less than 2℃ for daily maximum and minimum temperatures with 24 h lead time are increased from 46% and 66% before correcting to 63% and 74% after correcting in Shaanxi Province. These correction effects are higher than accuracies of the temperature forecast corrected by using air temperature data from observation and unrevised CLDAS.