Using a Hybrid Deep Learning and Statistical Approach to Enhance Subseasonal-to-Seasonal Precipitation Forecasts for Coastal Compound Flooding

An ongoing project is investigating the effects of seasonal to sub-seasonal (S2S) time scales (i.e., a time window spanning 1-28 days into the future) on compound flooding and inundation in coastal regions. One aspect of this project is to develop and apply bias-corrected ensemble precipitation forecasts from the North America Multi-Model Ensemble Phase 2 (NMME2) in a coupled coastal modeling system to quantify the impact at the S2S scales. Dynamical precipitation forecasts at a S2S could provide important information for critical societal functions. However, the coarse native spatial resolution, as well as the substantial forecast bias and error of S2S precipitation forecasts, limits its broader applications. Statistical techniques, such as Bias Correction and Spatial Disaggregation (BCSD), are commonly applied to post-process available dynamical precipitation forecasts for bias removal and spatial downscaling. However, BCSD is rather limited in improving the predictive skill of dynamical precipitation forecasts. In this study, we develop a hybrid framework that merges a popular Deep Learning (DL) technique, termed U-Net, with BCSD to post-process and downscale the daily S2S precipitation forecast from the coupled Canadian Climate Centre for Modelling and Analysis global climate model (CanCM3) over a study period from 1982 to 2011. Specifically, in the proposed framework, U-Net is first employed to incorporate multiple forecast variables, such as temperature and geopotential heights, to improve the predictive skill of the S2S precipitation forecast at its native spatial resolution. Then, the improved S2S precipitation forecast is passed to BCSD for bias removal and spatial disaggregation to 0.1-degree resolution. The performance of the proposed framework is tested over five study regions across the contiguous United States to consider both geospatial and climatological variabilities. The results suggest the proposed framework provides an equivalent level of bias removal to that of BCSD; more importantly, the proposed framework improves the predictive skill of the S2S precipitation forecast due to the employment of U-Net, especially in weeks one and four. Overall, our results highlight a possible protocol that could be applied to other advanced DL approaches to downscale dynamical forecasts for more accurate and reliable predictions of precipitation at S2S timescales