Evaluation of Multivariate Downscaling of GEOS S2S Meteorological Forecasts Using Climate Regionalization for the Amazon

With record-breaking heat and extreme precipitation events intensifying each year, accurate subseasonal to seasonal (S2S) forecasts assume prime importance in preparing for and managing these frequent extreme events. S2S forecasts typically rely on low-resolution global models for predictions, often insensitive to local variation of weather anomalies and their interactions with regional surface properties. This suggests the necessity of downscaling these forecasts to enhance their ability in capturing local and regional spatio-temporal patterns. This study presents an evaluation of downscaled Goddard Earth Observing System (GEOS) S2S meteorological hindcasts using the Generalized Analog Regression Downscaling (GARD) approach. Precipitation and 2m air temperature fields were downscaled from 0.5° to 0.125°, utilizing the Global Data Assimilation System (GDAS) as the observed data source. The approach employs multivariate downscaling using nine meteorological variables to effectively capture the local dynamics and global teleconnections. Additionally, climate regionalization is applied to classify regions exhibiting localized discrepancies. The forecast skill of the downscaled hindcasts is subsequently assessed for improvements in skill and bias reduction, particularly in regions identified with low skill metrics through climate regionalization.