Abstract

The accurate representation of spatio-temporal patterns of precipitation is an essential input for numerous environmental applications. However, the estimation of precipitation patterns derived solely from rain gauges is subject to large uncertainties. We present the Random Forest based MErging Procedure (RF-MEP), which combines information from ground-based measurements, state-of-the-art precipitation products, and topography-related features to improve the representation of the spatio-temporal distribution of precipitation, especially in data-scarce regions. RF-MEP is applied over Chile for 2000-2016, using daily measurements from 258 rain gauges for model training and 111 stations for validation. Two merged datasets were computed: RF-MEP3P (based on PERSIANN-CDR, ERA-Interim, and CHIRPSv2) and RF-MEP5P (which additionally includes CMORPHv1 and TRMM 3B42v7). The performances of the two merged products and those used in their computation were compared against MSWEPv2.2, which is a state-of-the-art global merged product. A validation using ground-based measurements was applied at different temporal scales using both continuous and categorical indices of performance. RF-MEP3P and RF-MEP5P outperformed all the precipitation datasets used in their computation, the products derived using other merging techniques, and generally outperformed MSWEPv2.2. The merged P products showed improvements in the linear correlation, bias, and variability of precipitation at different temporal scales, as well as in the probability of detection, the false alarm ratio, the frequency bias, and the critical success index for different precipitation intensities. RF-MEP performed well even when the training dataset was reduced to 10% of the available rain gauges. Our results suggest that RF-MEP could be successfully applied to any other region and to correct other climatological variables, assuming that ground-based data are available.