Understanding the dependence of the uncertainty in a satellite precipitation data set on the underlying surface and a correction method based on geographically weighted regression

It is widely acknowledged that the complicated underlying surface is one of the prominent reasons leading to serious uncertainty in satellite precipitation data sets over mountainous regions. However, no analysis has been conducted to quantitatively investigate the correlation between the errors in satellite precipitation data sets and the underlying surface. Using 133 monthly rain gauge observations over the Tibetan Plateau, the Bias and the residuals of ordinary least regression (the latter called ‘Ɛ’) in Climate Prediction Center morphing (CMORPH) data were calculated and were fitted with underlying surface factors using the geographically weighted regression (GWR) method, aiming at quantitatively understanding the dependence of the uncertainty in the CMORPH data set on the underlying surface over the Tibetan Plateau. We found that the 39.4% and 50.5% of the variance of the Bias and Ɛ, respectively, could be explained by the digital elevation model, the normalized difference vegetation index and land surface temperature. Furthermore, the explained variance of the Bias and Ɛ could be increased to 53.6% and 75.1%, respectively, by adding the CMORPH to the explanatory variables. Subsequently, the errors in the CMORPH were estimated by the GWR model, which was selected by comparing the explanation strengths of these models, and then the simulated errors were used to correct the precipitation estimated by CMORPH over areas without gauges. Independent validation indicated that the corrected CMORPH showed obviously better performance compared with the uncorrected CMORPH as well as two widely used precipitation estimates – the Universal Kriging interpolation model and the Tropical Rainfall Measuring Mission 3B43. These results reveal the significant influence of the underlying surface on the uncertainty in satellite precipitation data sets over mountainous areas and provide a promising approach to improve the precipitation estimates derived from satellite observations using underlying surface information.