Characterization and Modeling of Satellite-Based Precipitation Uncertainty over Complex Terrain

Difficulties in representation of high rainfall variability over mountainous areas using ground-based sensors is an open problem in hydrometeorological applications. These difficulties highlight the need to use integrated satellite-based precipitation products (SPP) because of their ability to represent the space-time variability of rainfall with quasi-global coverage. The rational and effective use of SPPs requires a thorough understanding of the error sources of the individual passive microwave (PMW) retrievals associated with each sensor used in the integrated products. Evaluation of PMW retrievals is challenging since it requires reference datasets with high temporal and spatial resolution. We address this PMW error assessment over complex terrain using dual polarization X-band radars. In this study we use SCOP-ME, which is an algorithm that uses best-fitted functions of specific attenuation coefficients and backscattering differential phase shift to retrieve rainfall rates and its microphysical characteristics from X-band dual-polarization radar observations. We are basing this study on X-band radar deployments in Northeast Italian Alps, North Carolina, Olympic Mountain, Southern tip of Vancouver Island, Rocky Mountains Colorado, Swiss Alps and Arizona. In situ observations from each site were used to evaluate the error characteristics of SCOP-ME retrieval and provide high-resolution estimates of the 4D rainfall variability in those mountainous areas. These estimates represent the benchmark precipitation dataset that is used in the error characterization and modeling of the PMW retrievals, namely GPROFV05 algorithm for MHS, SSMIS, GMI and AMSR2 sensors. For the error modeling we use the nonparametric tree-based quantile regression forests (QRF) model. The ensembles generated using the QRF model are validated by independent matchups of PMW/X-band precipitation data from the different study areas. The study shows that the error model significantly reduces both mean relative error and the random component of the error compared to the original PMW products. Moreover, the study shows that there is transferability of this error model between complex terrain regions, which will allow algorithm developers to integrate this error model to produce Level-3 products.