Temporal and spatial evaluation of satellite rainfall estimates over different regions in Latin-America

In developing countries, an accurate representation of the spatio-temporal variability of rainfall is currently severely limited, therefore, satellite-based rainfall estimates (SREs) are promising alternatives. In this work, six state-of-the-art SREs (TRMM 3B42v7, TRMM 3B42RT, CHIRPSv2, CMORPHv1, PERSIANN-CDR, and MSWEPv2) are evaluated over three different basins in Latin-America, using a point-to-pixel comparison at daily, monthly, and seasonal timescales. Three continuous (root mean squared error, modified Kling-Gupta efficiency, and percent bias) and three categorical (probability of detection, false alarm ratio, and frequency bias) indices are used to evaluate the performance of the different SREs, and to assess if the upscaling procedure used, in CHIRPSv2 and MSWEPv2, to enable a consistent point-to-pixel comparison affects the evaluation of the SREs performance at different time scales. Our results show that for Paraiba do Sul in Brazil, MSWEPv2 presented the best performance at daily and monthly time scales, while CHIRPSv2 performed the best at these timescales over the Magdalena River Basin in Colombia. In the Imperial River Basin in Chile, MSWEPv2 and CHIRPSv2 performed the best at daily and monthly time scales, respectively. When the basins were evaluated at seasonal scale, CMORPHv1 performed the best for DJF and SON, TRMM 3B42v7 for MAM, and PERSIANN-CDR for JJA over Imperial Basin. MSWEPv2 performed the best over Paraiba do Sul Basin for all seasons and CHIRPSv2 showed the best performance over Magdalena Basin. The Modified Kling-Gupta efficiency (KGE') proved to be a useful evaluation index because it decomposes the performance of the SREs into linear correlation, bias, and variability parameters, while the Root Mean Squared Error (RMSE) is not recommended for evaluating SREs performance because it gives more weight to high rainfall events and its results are not comparable between areas with different precipitation regimes. On the other hand, CHIRPSv2 and MSWEPv2 presented different performance, for some study areas and time scales, when evaluated with their original spatial resolution (0.05 degrees and 0.1, respectively) with respect to the evaluation resulting after applying the spatial upscaling (to a unified 0.25), showing that the upscaling procedure might impact the SRE performance. We finally conclude that a site-specific validation is needed before using any SRE, and we recommend to evaluate the SRE performance before and after applying any upscaling procedure in order to select the SRE that best represents the spatio-temporal precipitation patterns of a site.