Extreme value metastatistical analysis of remotely sensed rainfall in ungauged areas: Spatial downscaling and error modelling

The quantitative validation of rainfall statistics obtained from space-borne sensors, and the evaluation of the associated uncertainty are hindered by i) the limited coverage of rain gauge networks and weather radars at the ground, and ii) the different intrinsic statistical properties of point ground measurements and of area-averaged remote sensing estimates. This problem is particularly significant for extreme rainfall frequency analysis, where the inference process is heavily impacted by observational uncertainty and short records of satellite quantitative rainfall estimates (QPEs). Here we develop an approach to the validation and correction of QPE extreme rainfall statistics over data-scarce regions. We employ a statistical technique for downscaling QPE probability density function, spatial correlation structure, and frequency of extreme events to sub grid scales, so as to permit direct comparison with rain gauge point measurements. Extreme value modelling is performed using the Metastatistical Extreme Value (MEV) Distribution, which allows for the use of all available information from short QPE samples. We test this framework over the Conterminous United States (CONUS), providing a spatially-extended characterization of the performance of the Tropical Rainfall Measuring Mission (TRMM) Multisatellite Precipitation Analysis (TMPA) dataset. We then leverage this spatially-explicit analysis to develop a non-parametric model of the high-quantile estimation error. The model allows for the extrapolation of the QPE validation to ungauged locations. We test this methodology by means of a cross validation approach using independent observations at the ground over the CONUS. The results support the use of this approach to extreme rainfall estimation in ungauged areas, paving the way to similar applications at the global scale.