slide 1 model assimilation of satellite observations and retrievals andrew collard...

Model Assimilation of Satellite Observations and Retrievals

Andrew Collard

IMSG@NOAA/NCEP/EMC

Talk Outline

Introduction

The Global Model Cycle

Data Assimilation

Evaluating Impact

Data Impact Results

GOES-R and JPSS

Summary

Introduction

This talk will summarize how we make use of satellite data in operational forecast models

We will discuss how satellite observations and derived products are used to produce our best estimate of the current state of the atmosphere

This can be considered a data fusion process, where the information from current and historical observations are combined in an optimal manner.

The global model cycle

NOAA Satellite Proving Ground Slide 4

Data Assimilation


Data Assimilation

Data assimilation for NWP updates the atmospheric state which is used as a starting point for the forecast run. This state is the analysis.

It is a retrieval of the entire atmospheric state using multiple data sources simultaneously.

It is also, therefore, an efficient data fusion process.

The a priori or first-guess data used in the data assimilation process is a short range forecast from a previous cycle.

- The first guess, and therefore the analysis, contain information from all previous observations, propagated forward using the atmospheric model.


Atmospheric Analysis ProblemJ = Jb + Jo + Jc

J = (x-xb)TBx-1(x-xb) + (y-K(x))T(E+F)-1(y-K(x)) + JC

J = Fit to background + Fit to observations + constraints

x = Analysis

xb = Background (usually a short-range forecast from the previous cycle)

Bx = Background error covarianceK = Forward model (nonlinear)O = ObservationsE+F = R = Instrument error + Representativeness error

JC = Constraint term

The ensembles are run to give the data assimilation system, flow-dependent background errors

Atmospheric Analysis ProblemJ = Jb + Jo + Jc

J = (x-xb)TBx-1(x-xb) + (y-K(x))T(E+F)-1(y-K(x)) + JC

J = Fit to background + Fit to observations + constraints

x = Analysis

xb = Background (usually a short-range forecast from the previous cycle)

Bx = Background error covarianceK = Forward model (nonlinear)O = ObservationsE+F = R = Instrument error + Representativeness error

JC = Constraint term

Radiance Data Assimilated at NCEP

GOES-15 Sounders Channels 1-15

Meteosat-10 SEVIRI Channels 5-6

AMSU-A

NOAA-15 Channels 1-10, 12-13, 15

NOAA-18 Channels 1-8, 10-13, 15

NOAA-19 Channels 1-7, 9-13, 15

METOP-A Channels 1-6, 8-13, 15

METOP-B Channels 1-13, 15

AQUA Channels 6, 8-13

MHS

NOAA-18 Channels 1-5

NOAA-19 Channels 1-5

METOP-A Channels 1-5

METOP-B Channels 1-5


SSMIS

DMSP-17 Channels 1-7, 24

DMSP-18 Channel 1-7, 24

HIRS METOP-A Channels 2-15

AIRS AQUA 120 Channels

IASI

METOP-A 165 Channels

METOP-B 165 Channels

CrIS SNPP 84 Channels

Derived Products Assimilated at NCEP

Atmospheric Motion Vectors

- MODIS Aqua and Terra

- GOES-E, GOES-W

- MSG

- MTSAT

- VIIRS (pre-operational)

Ozone

- NOAA-19 SBUV Profiles

- OMI Total Column

GPSRO


Global Satellite Observing System

Data assimilation requires

Accurate and fast observation operators

- For radiances we use CRTMQuality control

- To handle situations where the observation cannot be used

Bias-correctionData quality monitoring


Application of NWP Bias Correction for SSMIS F18

Ascending Node Descending Node

Latitude

Unbias & Bias Corrected O-B

O-B Before Bias Correction

Global

Dsc

Asc

O-B After Bias Correction

Global

Dsc

Asc

O-B Before Bias Correction

O-B After Bias Correction

Using Met Office SSMIS Bias Correction Predictors

T (K)

T (K)

T (

K)

17

Quality Monitoring of Satellite Data

AIRS Channel 453 26 March 2007

Increase in SDFits to Guess

Data assimilation is statistical


A significant positive impact at day 8 still has a number of cases with negative impact.

Data assimilation is statistical (2)


Histogram of the individualimpact of each NOAA-19AMSU-A Ch 8 field-of-view

Verification


Verification

• Obviously the ultimate purpose of assimilating any kind of observation into NWP models is to improve the model forecast.

• Therefore robust methods for determining the impact of the observations are required. Operational centers typically have a wide range of tools available for this.

• The initial aim in any data assimilation system is to use the observations in such a way as to improve the accuracy in the analysis. Sometimes these improvements are not directly translated into increases in forecast skill due to the characteristics of the forecast model itself.

• In general this presentation is considering aircraft observations of temperature, humidity and wind vector together.

• But first, we need to define what we can use as the “truth” when verifying our models…

Slide 22

What is truth?

Observations?

Normally this means radiosondes (but also surface observations and aircraft are used). This means the statistics are biased towards densely populated regions in the northern temperate latitudes.

We should also (but usually don’t) take account of the errors in the observations themselves

Slide 23

Potentially we could also use satellite observations (e.g., radiances) for verificationas this would give more global sampling.However, comparisons in radiance spacewould be less intuitive.

What is truth?

Analyses?

The analysis should the best estimate of the atmospheric state through the combination of the information from the observations in the current and (through the forecast model) previous model cycles.

Also, given that the analysis does not suffer from the spatial sampling issues of conventional observations, the analysis seems to be the ideal “truth”.

However, for forecast ranges of approximately three days or less it is found that forecast skill (and even the sign of that skill) is highly dependent on the verifying analysis (control, test, independent) used.

Also, for certain types of changes, if additional structure is added to the analysis fields this can be penalized in the usual forecast skill measures as it is easier to obtain a good fit to smooth rather than complex fields.

Slide 24

What is truth?

In practice, both observations and analyses are used in verifying forecast skill

The 2014 NOAA Aircraft Workshop, ARINC Slide 25

Data Denial

Data denial or Observation System Experiments (OSEs) are simply a way of

investigating the impact of an observation or change by running full forecast

experiments with and without the element to be tested.

OSEs are expensive to run, particularly at full operational resolution, and they

need to be run for many forecast cycles (60 days is a typical number for global

forecast systems) before statistically significant results are obtained.

Individual case studies are generally not trusted as a way of demonstrating

forecast impact because of the dominance of statistical fluctuations.

Forecast impact scores are generally presented with error bars indicating

statistical significance.

Scores are normally given in terms of differences between forecasts and

“truth” in terms of RMS error or anomaly correlation coefficients (see next

slide)


Anomaly correlation coefficients


Forecast skill is often given in terms of anomaly correlation coefficients (ACC scores). This is a measure of the correlation between the forecast (f) and the analysis (a) while removing the climatological average field (c) to mitigate the seasonal signal:

Scores of 100% are perfect forecasts, 60% is considered the lower limit for useful skill.

xb

xg

t= -6 hrs

eg

xt

xf

xa

t= 24 hrst=0

ef

6 hr assimilation window

Observations move the forecast from the background trajectory to the trajectory starting from the new

analysis

In this context, “OBSERVATION IMPACT” is the effect of observations on the difference in forecast error norms C(ef-eg)

Forecast Sensitivity to Observations (FSO)

The 2014 NOAA Aircraft Workshop, ARINC 28

“Truth”“Background”

“Analysis”

30-hr fcst

24-hr fcst

Langland and Baker (Tellus, 2004), Gelaro et al (2007), Morneau et al. (2006)

xb

xg

t= -6 hrs

eg

xt

xf

xa

t= 24 hrst=0

ef

6 hr assimilation window

Observations move the forecast from the background trajectory to the trajectory starting from the new

analysis

In this context, “OBSERVATION IMPACT” is the effect of observations on the difference in forecast error norms C(ef-eg)


The 2014 NOAA Aircraft Workshop, ARINC 29

“Truth”“Background”

“Analysis”

30-hr fcst

24-hr fcst

(ef-eg)=Mkδy

δy = observation innovations (y-H(xb))

K=Kalman gain (transforms observation innovations to analysis increments)

M=Forecast model (transforms analysis to forecast)

Langland and Baker (Tellus, 2004), Gelaro et al (2007), Morneau et al. (2006)

Forecast Sensitivity to Observations (FSO)We want to get the sensitivity of the forecast to the observation increments so we apply the tangent linear model to C(ef-eg)=CMKδy:

δef-g = ½[MKδy]TC(ef-eg) = ½ δyTMTKTC(ef-eg)


Forecast Sensitivity to Observations (FSO)We want to get the sensitivity of the forecast to the observation increments so we apply the tangent linear model to C(ef-eg)=CMKδy:

δef-g = ½[MKδy]TC(ef-eg) = ½ δyTMTKTC(ef-eg)


Adjoint of data assimilation scheme

Adjoint of linearized forecast model*

Usually these both require approximations to be made (including linearity) and so this method is limited to forecast ranges of less than ~48 hours.

*Already required if running 4DVar

Advantages and Disadvantages of FSO• Advantages

• Can infer the impact of observations to whatever level of detail is required (e.g. ob by ob, channel by channel) without having to re-run the full system repeatedly.• Useful for determining relative impact of observations and

for quality control of bad observations.• Allows the impact of observations on the forecast to be

monitored on a daily basis.• Disadvantages

• Limited to short-range forecasts • So there is sensitivity to the accuracy of the verifying

analysis• Impact is always in the context of the total observing system as

used• Forecast impacts of an observation type may change as

other observations are added/removed.


Data denial experiments

At NCEP we are still developing an FSO capability based on the EnKF methodology.

So we will present an overview of observation impact based on the data denial method.


Data Denial Experiments(Jim Jung and Lars-Peter Riisshojgaard)


Background

NCEP Operational GDAS/GFS May 2011 versionT574L64 operational resolutionTwo Seasons

- Aug-Sept 2010

- Dec 2010-Jan 2011

Cycled experiments 7 Day forecast at 00ZControl late analysis (GDAS) used for

verificationNot NCEP operations computer

36

500 hPa Anomaly Correlations15 Aug – 30 Sep 2010

37

No Satellite / No Conventional Data

Northern Hemisphere Southern Hemisphere

500 hPa Anomaly Correlations 15 Aug – 30 Sep 2010

38

No AMSU-A / No MHS



39

No GPS-RO / No AMV



40

No Rawinsondes / No Aircraft



41

No Hyperspectral Infrared


500 hPa, Day 5, Instrument Average AC scores

42

Data Addition Experiments(Jim Jung and Mitch Goldberg)


Background

44 April 21, 2023

• Pre 2015 version of the GDAS/GFS Hybrid (80 ensembles, T254) T670 Semi-Lagrangian

• Winter and summer seasons 2014• Baseline is conventional data and GPS-RO• Add single instruments

ATMS (SNPP), CrIS (SNPP) AMSUA, MHS (Metop-b), IASI (Metop-a) SSMIS (F18), AIRS (Aqua)

• Verified against the late control analysis with all operational data

250 hPa Northern Hemisphere AC scores for 20140801 – 20140831 00Z

45 April 21, 2023

250 hPa Southern Hemisphere AC scores for 20140801 – 20140831 00Z

46 April 21, 2023


47 April 21, 2023


48 April 21, 2023


49 April 21, 2023


50 April 21, 2023

200 hPa Tropical Vector Wind RMSE for 20140801 – 20140831 00Z

51 April 21, 2023

850 hPa Tropical Vector Wind RMSE for 20140801 – 20140831 00Z

52 April 21, 2023

JPSS and GOES-R


So how will JPSS/GOES-R fit in?ATMS and CrIS will be assimilated closely following our

experience with the NPP instruments

- We will continue to move towards greater use of cloudy radiances

- We will make more aggressive use of CrIS channels – both in terms of observation error and the number used.

VIIRS will be used primarily through the winds and SST analysis.

OMPS ozone

GOES-R will provide an initial impact through the AMVs

There are significant challenges still with using the full GOES-R radiance data stream, so averaged products will be useful.


Summary


Summary

Data assimilation is an efficient form of data fusion.

The global observing system combined with the data assimilation system is very robust.

- Individual data types have an incremental impact on forecast skill but the combined effect is very large.

Data addition experiments can give greater insight into the impact of each individual data type

We can expect rapid implementation (dependent on model upgrade implementation cycles) of CrIS, ATMS, VIIRS AMVs and GOES-R AMVs based on current experience with existing data.


slide 1 model assimilation of satellite observations and retrievals andrew collard...

Documents