Ensemble-based Stochastic Downscaling of Seasonal Predictions of Extreme Rainfall (invited)

There is increasing interest in extreme weather and climate events, both for the development of weather early warning systems and to better understand the impacts of climate change on fine-scale weather events. End users of S2S forecasts need not only skillful forecasts of rainfall totals but also of extremes. Most extreme rainfall events are due to strong convective processes with horizontal scales only a few kilometers, which cannot be directly resolved by current S2S models. Nevertheless, S2S models give reliable information on the seasonal features of the atmospheric circulation that encompass the subgrid-scale processes leading to extremes.

To date, the extent to which statistical postprocessing can improve forecasts of rainfall extremes from current S2S models has not been extensively explored. This study will use StormLab, (Liu et al. 2024) a data-driven stochastic downscaling technique, to compare downscaled S2S rainfall forecasts (NASA’S GEOS-S2S-v3) against a satellite-based precipitation (SPP) product (NASA’s IMERG). StormLab addresses several major shortcomings of previous stochastic rainfall synthesis approaches, including by providing a robust relationship between large-scale atmospheric conditions and fine-scale precipitation. StormLab has two key components: (i) a parametric statistical uncertainty model to describe the relationships between large-scale atmospheric variables and fine-scale rainfall intensities; and (ii) STREAM, an autocorrelation modeling method to generate realistic non-stationary stochastic rainfall ensembles (Hartke et al. 2022). The relationship between large-scale atmospheric variables (integrated water vapor (IVT) in this case, calculated from NASA’s GiOCEAN) and fine-scale rainfall magnitudes is modeled by a parametric generalized Gamma-based statistical model. Next, a “localized” short-space Fourier transform and meta-Gaussian approach is used to rapidly sample from the Gamma-based statistical model to generate rainfall fields that have realistic structure (spatiotemporal autocorrelation, anisotropy, spatial nonstationarity) consistent with both the large-scale and fine-scale fields.

We will present on recent extreme precipitation events and validate the skill of downscaled S2S precipitation forecast ensembles. This work underscores the potential of stochastic downscaling methods in enhancing the predictive capabilities of S2S models, thereby contributing to more effective water resource management and disaster preparedness strategies.