Correction of TRMM Monthly Precipitation Data from 1998 to 2013 in Xinjiang

Using monthly TRMM precipitation data and precipitation observations from 105 national basic weather stations in Xinjiang region from 1998 to 2013, a stepwise regression model and a back-propagation (BP) neural network are established to correct TRMM precipitation. Results show that models added with geographical factors can increase the accuracy of TRMM precipitation effectively. Corrected by two models, overall correlation coefficients are 0.75 and 0.80, and relative errors are 4.88% and 3.19%. On the monthly scale, the relative error of TRMM monthly precipitation ranges from-5.68% to 54.44%, from-4.26% to 32.57% after stepwise regression and from 5.33% to 24.48% after neural network, respectively. In addition, results show that qualities of satellite precipitation products are improved in different degrees from S_T, with 0.01-0.49 for stepwise regression model and 0.03-0.70 for neural network, respectively. Compared with TRMM data before correction, the stepwise regression model and BP neural network model can accurately and quantitatively reproduce the actual distribution of precipitation, and provide a more practical method for the area lack of precipitation data.