Precipitation Merging Based on the Triple Collocation method across Mainland China

Triple collocation (TC) is a novel method for quantifying the uncertainties of three independent inputs and has been widely used for various geographical variables such as precipitation, wind speed, soil moisture, sea ice thickness. TC shows potential in the merging of multiple inputs, which however has been not used to precipitation. Therefore, based on the TC formulation, this study merges precipitation from the Climate Prediction Center's morphing technique (CMORPH), Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN) and the 5-th generation European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA-5). Ground observations from more than 2000 rain gauges in China are used as the validation benchmark. Two strategies (merging rainfall and snowfall separately and merging precipitation directly) were designed because one product could show different accuracies concerning rainfall and snowfall component. In addition, two weighting method using Root Mean Square Error (RMSE) in logarithmic scale (log-RMSE) and modified scale (mod-RMSE) are compared. Results show that firstly the TC-based merging method is effective in precipitation, although it is susceptible to the underestimation of inputs. Besides, merging rainfall and snowfall separately is superior to merging precipitation directly. The improvement is particularly notable in winter. Last but not least, the mod-RMSE shows worse performance in weight estimation than log-RMSE due to the underestimation of inputs. This study not only proves the promising potential of TC but highlights the importance of separate treatment of rainfall and snowfall in precipitation merging. However, there are two limitations in the application of TC: (1) only three inputs are allowed and (2) the three inputs must be uncorrelated with each other. To overcome the deficiencies, we are conducting research to generalize TC by combining the least square method and iteration for first and dividing error within each input product into two uncorrelated components for second.