Open AccessStatistical Downscaling Multimodel Forecasts for Seasonal Precipitation and Surface Temperature over the Southeastern United States
Author(s) -
Di Tian,
Christopher J. Martinez,
Wendy D. Graham,
Syewoon Hwang
Publication year - 2014
Publication title -
journal of climate
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 3.315
H-Index - 287
eISSN - 1520-0442
pISSN - 0894-8755
DOI - 10.1175/jcli-d-13-00481.1
Subject(s) - downscaling , climatology , forecast skill , environmental science , precipitation , quantile , climate forecast system , meteorology , sea surface temperature , climate model , coupled model intercomparison project , ensemble average , statistics , mathematics , climate change , geography , ecology , biology , geology
This study compared two types of approaches to downscale seasonal precipitation (P) and 2-m air temperature (T2M) forecasts from the North American Multimodel Ensemble (NMME) over the states of Alabama, Georgia, and Florida in the southeastern United States (SEUS). Each NMME model forecast was evaluated. Two multimodel ensemble (MME) schemes were tested by assigning equal weight to all forecast members (SuperEns) or by assigning equal weights to each model’s ensemble mean (MeanEns). One type of downscaling approach used was a model output statistics (MOS) method, which was based on direct spatial disaggregation and bias correction of the NMME P and T2M forecasts using the quantile mapping technique [spatial disaggregation with bias correction (SDBC)]. The other type of approach used was a perfect prognosis (PP) approach using nonparametric locally weighted polynomial regression (LWPR) models, which used the NMME forecasts of Niño-3.4 sea surface temperatures (SSTs) to predict local-scale P and T2M. Both SDBC and LWPR downscaled P showed skill in winter but no skill or limited skill in summer at all lead times for all NMME models. The SDBC downscaled T2M were skillful only for the Climate Forecast System, version 2 (CFSv2), model even at far lead times, whereas the LWPR downscaled T2M showed limited skill or no skill for all NMME models. In many cases, the LWPR method showed significantly higher skill than the SDBC. After bias correction, the SuperEns mostly showed higher skill than the MeanEns and most of the single models, but its skill did not outperform the best single model.