Young-Kwon Lim, D.W. Shin, S. Cocke, T. E. LaRow, J. J. O’Brien, and E. P. Chassignet Center for Ocean-Atmospheric Prediction Studies, Florida State University, Tallahassee, FL, USA Statistical Downscaling of the NCEP CFS Retrospective forecasts (precipitation) over the SE US
27
Embed
Statistical Downscaling of the NCEP CFS Retrospective forecasts (precipitation) over the SE US
Statistical Downscaling of the NCEP CFS Retrospective forecasts (precipitation) over the SE US. Young-Kwon Lim, D.W. Shin, S. Cocke, T. E. LaRow, J. J. O’Brien, and E. P. Chassignet Center for Ocean-Atmospheric Prediction Studies, Florida State University, Tallahassee, FL, USA. - PowerPoint PPT Presentation
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Young-Kwon Lim, D.W. Shin, S. Cocke, T. E. LaRow, J. J. O’Brien, and E. P. Chassignet
Center for Ocean-Atmospheric Prediction Studies, Florida State University, Tallahassee, FL, USA
Statistical Downscaling of the NCEP CFS Retrospective forecasts (precipitation)
over the SE US
Background and Motivation
Global NCEP/CFS : 1) Retrospective forecasts longer than 20 year period (1981-2006), 2) Widely used in many studies, 3) the low seasonal predictive skill (e.g., precipitation for growing season) in certain areas.
Question: Can we successfully downscale the CFS data which have 2.5 degree resolution and the low skill over several regions?
Why downscaling over the SE USA?
Extremely high temperature and heavy rainfall with severe storms during summer, resulting in potential property damage and injuries.
The largest areas of agricultural farms in the nation.
An accurate forecast with higher spatial resolution is essential to adapt management, increase profits, reduce production risks, and mitigate damages.
Regional climate simulation in FSU/COAPS
FSU/COAPS Global Spectral Model (FSU/COAPS GSM) has been downscaled to the 20km grid resolution by FSU/COAPS nested regional spectral model (FSU/COAPS NRSM) over the southeast US. Dynamical Downscaling
Statistical downscaling model has been also developed. (CSEOF, multiple regression, and stochastic PC generation are used.)
Training Predictor : model output
Predictand : observation
&
Regressed eigenfunctions of CFS runs used
0.2° 0.2° (~20km res.) 2.5° 2.5° (~250km res.)
Eigenfunctions of the Obs. over training period and the Generated PC used
Prediction period
Withholding different Withholding different year for year for Cross-Cross-validationvalidation
Data (Obs. & CFS) and period
Variables : Daily precipitation
Period : 1987 ~ 2005 (Spring (MAM) ~ Summer (JJA) each year (daily))
Observed data source :
National Weather Service Cooperative Observing Program surface data over the southeast US : ~20km×20km
Large-scale data to be downscaled :
NCEP/CFS retrospecitve forecasts : 2.5°×2.5°, 10 members with lagged initial conditions. Seasonal integrations starting from February each year.
Results
2-d seasonal mean field (CFS, Downscaled data, and Observation)
Time series over ~20 years (Interannual variation) for three states
Precipitation for growing season from NCEP/CFS (~2.5° res.) run have been downscaled to local scale of ~20km for the SE US.
Downscaling simulates the regional-scale seasonal precipitation with reduction in wet biases.
Correlation, categorical predictability for seasonal anomaly has been improved from the coarsely resolved NCEP/CFS.
Heavy rainfall events: In overall, downscaling better produces the interannual frequency variation than bias-corrected rescaling.
Subseasonal dry spells: Rescaled data show significant underestimation with much smaller amplitude variation than observation.
Application to crop model and realtime forecast.
Statistical downscaling procedure (1)
1. Cyclostationary EOF analysis for the model output and the observation :
CSEOF (Kim and North 1997) : analysis technique for extracting the spatio-temporal evolution of physical modes (e.g., seasonal cycle, ENSO, ISOs, etc.) and their long-term amplitude variations.
P(r,t)=∑n Sn(t) Bn(r,t)
Bn(r,t) : time-dependent eigenfunctions, Sn(t) : PC time series. In this study, CSEOF is conducted on both observation and
FSUGSM runs over the training period.
Statistical downscaling procedure (2)
2. Multiple regression between the model output and the observation :
CSFOF PC time series of the first significant modes of a predictor variable (FSUGSM data) are regressed onto a certain PC time series of the target variable (observation) in the training period.
PCTn(t)=∑iαni·PCPi(t)+ε(t) i=1,2,…10
PCTn(t): target PC time series, αni: regression coefficient
PCPi(t): predictor PC time series
Relationship between model output and the observation is extracted from CSEOF and multiple regression.
3. Generating CSEOF PC of the model data over the forecast period from the regressed fields in the training :
CSFOF PC time series of the model data are generated for the prediction period. Modeled data and the regressed eigenfunctions identified from training are used.
PCn(t)=∑gP(g,t)·Bn+(g,t)
PCn(t): the nth mode PC time series for the prediction period g : large-scale grid point
Bn+(g,t) : regressed CSEOF eigenfunctions
P(g,t): global model anomaly over the prediction period
Statistical downscaling procedure (4)
4. Downscaled data construction from the eigenfunctions of the observation and the generated CSEOF PC time series :
D(s,t)=∑nPCn(t)·Bno(s,t)
PCn(t) : generated PC time series from the previous step
Bno(s,t): CSEOF eigenfunctions of the observation (training
period)
D(s,t) : downscaled output
5. Generating downscaled output for the entire period (9yrs) by cross-validation framework
Black : ObservationRed : Downscaling
Blue : Rescaling from the CFS
Observed variation is captured by downscaling to a certain extent.
Several peaks are not captured well (e.g., 1998 in Florida).
Rescaled data with bias-correction oscillates near zero (significant underestimation).
Extremes (Frequency of Subseasonal dry spells (anomaly))