How well do deep learning-based downscaling and convection permitting simulations represent temperature and precipitation individual and compound daily extremes over Southeastern South America?

BETTOLLI MARIA LAURA; OLMO, M.; BALMACEDA HUARTE, R; BAÑO MEDINA, J

Fort Collins

Taller; VIII Convection Permitting Climate Modeling Workshop; 2024

Colorado State University, the NSF National Center for Atmospheric Research, and GEWEX

Southeastern South America (SESA) is highly exposed to extreme precipitation events that are usually accompanied by severe weather that have large socio-economic impacts. The region is also the scene of temperature extremes that occurs individually (such as heat waves) or jointly with precipitation extremes becoming a compound extreme event. With a focus on extreme rainfall events in SESA, one of the objectives of the Flagship Pilot Study in Southeastern South America (FPS-SESA) initiative endorsed by CORDEX is to develop actionable climate information from statistical (ESD) and dynamical downscaling. To this end, targeted convection-permitting (CPRCM) covering 3 consecutive years from June 2018 to May 2021 were conducted. In addition, ESD simulations based on convolutional neural networks (CNNs) were performed using the 1979-2017 period as a training period and tested in the independent 3-year period selected. In this work, we evaluate and intercompare both downscaling approaches in reproducing individual and joint occurrences of daily temperature and precipitation extremes against CPC Global Unified Temperature and Precipitation Dataset and ERA5 reanalyses. Overall, CPRCM and CNNs show skillful performance in modeling daily precipitation and temperature individual extremes, with larger spread in the former but within the observational uncertainty. Regarding the compound extremes, although a low number of cases could be assessed, the models were able to reproduce their occurrence and main characteristics, such as the spatial variability and location. These results contribute to the identification of the strengths and weaknesses of both approaches in reproducing temperature and precipitation extremes over SESA, and settle the basis for the exploration of the combination of both approaches to leverage their potential in a future CPRCM emulation study based on machine learning techniques.