Uncertainties of Future Extreme Precipitation Changes in Statistical Downscaling Products: Comparison between BC_DLD and LOCA2 over the US Northeast

Statistical downscaling is a computationally efficient approach to producing a large ensemble of fine resolution future climate projections to support locally relevant climate adaptation efforts. However, extreme precipitation from statistical downscaling products is subject to large uncertainties, and often depends on both the observational reference data and the downscaling algorithm. In response to the need for locally actionable climate information in the Northeast, a region experiencing the fastest increase of heavy precipitation, we developed a deep learning-based climate downscaling framework with bias correction (BC_DLD) and applied it to the daily output from the historical and SSP5-8.5 future scenario runs of four CMIP6 ESMs. We derived the extreme precipitation metrics from our downscaled product, and compared them against those from the same ESM ensemble member in the Localized Constructive Analogs version 2 (LOCA2) product over the Northeast. The two agree well in both the spatial variability and future changes of some precipitation statistics, but differ substantially in several extreme precipitation metrics that have significant societal implications. For example, the BC_DLD produced a slightly higher 20-year return value of daily precipitation (Rx1_20) for the historical climate than LOCA2, but this difference is amplified in future projections. Specifically, BC_DLD suggested a significant increase of Rx1_20 across majority of the Northeast, while LOCA2 produced a mixed signal with no spatially coherent changes. Another metric we analyze is how well the statistically downscaled products retain the emergent relationship between extreme precipitation and temperature, an aspect that both products are expected to reproduce, for different reasons. Specifically, the analogs for both precipitation and temperature were from the same day in LOCA2, while BC_DLD trained the deep learning model for precipitation using temperature as a predictor. However, to what extent each of these products retains the precipitation-temperature relationship in the reference data is not clear. Through comparing the BC_DLD and LOCA2 products in the Northeast focusing on the extreme precipitation metrics and the relationships between different variables, this presentation will highlight the challenges facing statistical downscaling and propose potential solutions.