Slide 1 Evaluation of observation impact and observation error covariance retuning Cristina Lupu , Carla Cardinali, Tony McNally ECMWF, Reading, UK WWOSC 2014, Montréal, Canada
Dec 23, 2015
Slide 1
Evaluation of observation impact and
observation error covariance retuning
Cristina Lupu, Carla Cardinali, Tony McNally
ECMWF, Reading, UK
WWOSC 2014, Montréal, Canada
Slide 2
Outline
1. Motivation
2. Estimating observation error variances
3. Assimilation experiments with an updated diagonal R
4. Summary
WWOSC 2014, Montréal, Canada
Slide 3
Motivation• Current operational ECMWF system is quite complex: ~ 40 millions observations from 60 instruments are daily assimilated.
• The assumed R and B play an important role in determining the weight of a given observation in the assimilation system.
• The estimation of the error covariances remains a significant challenge.
• In assimilation systems, the observation error covariance R describes errors in the observations as well as the forward model;
• We assume a diagonal R.
• Diagnostics tools are used to quantify the impact of all observations in ECMWF system both in analysis and forecasts.
WWOSC 2014, Montréal, Canada
Slide 4
Ways to estimate R • Diagnostics based on output from DA systems:
• Desroziers method (Desroziers et al., 2005) An estimate of the observation error variances may be obtained a posteriori from the statistical analysis of the observation residuals.
• Adjoint-based methods: makes forecast sensitivity to data assimilation system input parameters [ y, R, xb , B] possible:
• Forecast sensitivity to observations (FSO) – is used to monitor the impact of observations to reduce short-range forecast errors.
• Forecast R-sensitivity (Daescu & Todling, 2010; Daescu & Langland, 2013)may be used to provide guidance to error covariance tuning procedures.The sensitivity of a scalar measure of forecast error is computed with respect to changes to a set of covariance parameters .
• … WWOSC 2014, Montréal, Canada
Slide 5
Initial assimilation experiment
• Aim: Investigate the benefits of an updated (diagonal) R compared to the operational assimilation of IASI/Metop-A
• Baseline experiment assimilating only conventional observations and IASI /Metop-A with R diagonal as in ECMWF operations.
• Setup: 4D-Var, T511 (~ 40 km resolution), 137 vertical levels; Period: 8 June – 30 July 2012
• The simplified (in terms of observation usage) experiment intends to provide the backbone process for observation error variances tuning.
WWOSC 2014, Montréal, Canada
Slide 6
Desroziers diagnostic for σo for IASI
• The observation error standard deviations (σo) assumed in our system are strongly inflated.
WWOSC 2014, Montréal, Canada
Slide 7
Adjoint-based methods
FSO R-sensitivity (FSR)
WWOSC 2014, Montréal, Canada
FSO: The impact of observation is beneficial in each analysis cycle and reduces 24-h forecast error over the global domain by an average of 23.06%; IASI and AIREP observations are contributing the most to 24-h forecast impact.
FSR: Positive sensitivities: identify those observation types whose error variance deflation (decreasing the σo ) is of potential benefit to the 24-h forecast;
Slide 8
FSO
IASI (σo)2 – sensitivity guidance
WWOSC 2014, Montréal, Canada
IASI channels:
Positive sensitivities: Long-wave CO2 temperature-sounding channels;
Negative sensitivities: O3 band (range 1479-1671 )
• Inflation of the assigned observation error σo is of potential benefit to the forecast
• An observation error σo specification according with Desroziers estimates may have a detrimental forecast impact.
FSO, ch. 1671, O3 band
Slide 9
Adjusting (σo)2 for selected IASI channels
WWOSC 2014, Montréal, Canada
Use of adjoint-methods for tuning of observation error involved two steps:
•Use FSR to identify/select IASI channels where observation error standard deviations (σo) should be decreased/increased.
•For selected channels, use Desroziers estimates of (σo) to quantify how much this should be changed.
Slide 10
Experiments with an updated R
• Aim: Investigate the benefits of an updated R compared to the operational assimilation of IASI/Metop-A
• Baseline: assimilating only conventional observations and IASI /Metop-A with R diagonal as in ECMWF operations.
• Exp.1: As Baseline, but with updated diagonal R for all IASI channels as derived from Desroziers diagnostic;
• Exp.2: As Baseline, but with updated diagonal R for 33 selected IASI channels (temperature-sounding channels 173-254 and WV channels 2889-5480).
WWOSC 2014, Montréal, Canada
Slide 11
IASI AIREP-T
Impact on FG-departures
WWOSC 2014, Montréal, Canada
Normalised by Baseline, 95% confidence interval
WIND-U,V
Exp.1Exp.2
Slide 12
Forecast scores: geopotential
WWOSC 2014, Montréal, Canada
Normalised change in RMS geopotential forecast error at 500 hPaVerified against operational analysis; 95% confidence error bars
54 days summer 2012.
Nor
mal
ized
diff
eren
ce
Exp.1 - Baseline
Exp.2 - Baseline
Better than Baseline
Worse than Baseline
Forecast dayForecast day
Slide 13
Total dry energy error norm
WWOSC 2014, Montréal, Canada
• The energy norm evaluates the entire model volume of the atmosphere and calculates a combined error from four meteorological variables.
N. Hem. S. Hem.
• Using the Desroziers diagnosed σo for all IASI channels results in a degradation of analysis and subsequent forecasts .
Slide 14
Analysis sensitivity to observations
WWOSC 2014, Montréal, Canada
Analysis sensitivity Baseline Exp.2
Global Observation influence 9.3% 10.3%
Background influence 90.7% 89.7%
Slide 15
Jo – statistics per channel: IASI
WWOSC 2014, Montréal, Canada
Current obs. errors(Baseline)
Based on effective departure: deff = (y – H(x))T R-½
New obs. errors(Exp.2)
Slide 16
Forecast R- and B-sensitivities
WWOSC 2014, Montréal, Canada
•Positive R-sensitivities for all observation types : decreasing σo for all obs.
•The B-sensitivity provides guidance on weighting in the assimilation system between the background state and the whole observing system: background error covariance inflation.
•An optimal weighting between B and R information may be explained through a single covariance weight coefficient.
Slide 17
Summary• Results of a study aimed at tuning observation errors variances for IASI/Metop-A based on two methods: a posteriori diagnosis and adjoint-based R-sensitivity.
• Using the Desroziers diagnosed σo for all IASI channels results in a degradation in analysis and subsequent forecasts .
• Forecast R-sensitivity: found to be promising for providing guidance on IASI channel selection, but does not provide the amount of how much the observation-error variances should be changed.
• Beneficial forecast impact of geopotential, wind, temperature over the operational R.
• Forecast R- and B-sensitivities can provide guidance toward the real covariance matrices. The method may show if background information is being over (or under) weighted. In this case it appears the EDA based background errors are overweighting the background.
WWOSC 2014, Montréal, Canada
Slide 18
Open issues • Using Forecast sensitivity to R (FSR) to tune R in the current operational ECMWF system is a challenge:
• very large number of assimilated observations• what modifications to R do we need and why?
• In the ECMWF system, an ensemble of data assimilations is used to specify background errors.• The assumed R is used in the ensemble to perturb observations.• Need to investigate the impact of the new R on background error estimate.
WWOSC 2014, Montréal, Canada
Slide 19
Thank you!Questions?
WWOSC 2014, Montréal, Canada