The weather has a profound influence on human’s daily life and the weather forecasting has always been a topic of great concern. With the economic development and social progress, people’s requirements for daily weather forecasting has become higher and higher. Information provided by the general circulation models (GCMs) can describe well some of the weather parameters at a large scale, but GCMs fail to provide detailed weather information at a regional or local scale for impact assessment studies. Outputs from GCMs are usually of low spatial resolutions. A common approach to bridge the scale mismatch is downscaling. In the present study, two methods, i.e., the statistical multiple linear regression and the BP neural network, were proposed to downscale large scale reanalysis data to daily temperature extremums at a local point, Shenzhen national meteorological station. The data used in this study are NCEP/NCAR (National Centers for Environmental Prediction/National Centre for Atmospheric Research) reanalysis dataset for the 2000~2012 period and daily observations of maximum temperature and minimum temperature at Shenzhen station for the same period. The two methods were compared in this study. Results show that both methods can simulate well the daily temperature extremums at Shenzhen station, but the performance of the statistical downscaling method is more stable than the BP neural network.