A Bayesian Approach to Blend Satellite Precipitation Products and Ground Gauge Measurements in the Yangtze River Basin

Precipitation is the key forcing factor that provides essential input information for estimating land surface hydrological fluxes and states. Conventional gauges have served as the primary instrument to monitor precipitation for a long time; however, they are often limited in representing the spatial pattern of precipitation due to their sparse or uneven distribution. An alternative method relying on satellite-estimated precipitation products has substantially enhanced our capabilities for large-scale hydrological monitoring and modeling, although the estimation bias always tends to be significant. In this study, we propose a Bayesian merging approach to blend ground gauges measurements and two sets of satellite-derived precipitation products (TMPA 3B42 RT and CMORPH) in the Yangtze River basin, China. First, two types of general error models (i.e. the additive error model, and the multiplicative error model) as well as their parameters are estimated and intercompared for 3B24 RT and CMORPH, respectively, which serves as the likelihood function during the final Bayesian derivation. Meanwhile, prior spatial distribution of precipitation is obtained by gauge observations applying an ordinary Kriging approach, and the interpolation results offer initial estimates of both the mean value and the variance of precipitation at any location. With statistical assumputions and near real-time satellite-derived precipitation estimator from 3B42 RT and CMORPH, the posterior estimates of precipitation and their uncertainties then can be analytically derived by integrating the interpolated prior fields and the fitted likelihood function. Lastly, this Bayesian merging approach has been applied in the Yangtze River basin for monthly and daily precipitation estimation during an eight-year period (2008-2015), and the approach’s performance and its relations with some key parameters (e.g. gauge density) are further investigated and discussed. The evaluation results show that the merging approach has significantly improves upon the two satellite-derived estimates, and it also provides a promising way to probabilistically estimate precipitation at any given location by assimilating available multi-source data.