Integrating Community RT Components into JCSDA CRTM

1
Integrating Community RT Components into JCSDA
CRTM

• Yong Han, Paul van Delst, Quanhua Liu, Fuzhong
  Weng, Thomas J. Kleespies, Larry M. McMillin

2
Outline

• Part I
• Project objective
• Approach
• CRTM components
• CRTM implementation status
• Plans
• Issues
• Part II
• CRTM framework (Paul van Delst)

JCSDA 3rd Workshop on Satellite Data
Assimilation, 20-21 April 2005
3
Project Objective
Fast and accurate community radiative transfer
model to enable assimilation of satellite
radiances under all weather conditions
4
Approach

• Integrate community RT components
• Provide CRTM framework to the community to
  minimize efforts in integrating RT components
  into the CRTM
• Interact with the community research groups
  during the integration process assisting
  implementation and modifying the framework to
  accommodate their needs.

5
CRTM Components
public interfaces
Forward CRTM
CRTM Initialization
CRTM Destruction
Jacobian CRTM
Surface Emissivity/Reflectivity Model(s)
Aerosol Absorption/Scattering Model
Gaseous Absorption Model
Cloud Absorption/Scattering Model
RT Solution
Source Functions
6
CRTM Framework

• By Nov. 2004, the framework for both forward and
  Jacobian models was completed and distributed
  together with the documents.
• The framework details user and developer
  interfaces, data structures and program layouts
• The community is now using the framework as a
  vehicle to integrate RT components into the CRTM

7
Gaseous Absorption Module

• Function provide gaseous (water vapor, Ozone,
  dry gases, etc.) optical depth profiles
• Models OPTRAN and OSS (AER)
• Integration status
• OPTRAN forward, Tangent-linear and Adjoint models
  have been integrated with the CRTM framework and
  tested.
• OSS forward model has been preliminarily
  integrated with the CRTM framework
• OSS- and OPTRAN-based CRTMs

8
OPTRAN-based CRTM flowchart
CRTM Initialization
Channel Loop
channel i
Gaseous Optical depth (OPTRAN)
OPTRAN transmittance coefficients
Cloud optical parameters
Cloud optical parameter lookup tables
Aerosol optical parameters
Aerosol optical parameter database
Surface emiss. reflect.
Surface emissivity and reflectivity database
RT Solution
Computer memory
R_chi
no
Channel loop done?
yes
R_ch1 , R_ch2, , R_chn
9
OSS-based CRTM flowchart
CRTM Initialization
Node Loop
Node i
Gaseous Optical depth (OSS)
OSS OD lookup table
Cloud optical parameters
Cloud optical parameter lookup tables
Loop over those channels engaged with node i
Aerosol optical parameter database
Aerosol optical parameters
R_chk R_chk wkRi
Surface emiss. reflect.
Surface emissivity and reflectivity database
no
Channel loop done?
yes
RT Solution
no
Computer memory
Node loop done?
yes
R_ch1 , R_ch2, , R_chn
OSS weights node-channel map
Computer memory
10
Surface Emissivity Reflectivity Models
Microwave Land LandEM (Weng et al., 2001)
Snow and sea ice (Yan Weng, 2003) Ocean
wind vector dependent (Liu and Weng, 2003) wind
speed dependent (English,
1998) Infrared Ocean IRSSE (van Delst,
2003 Wu-Smith, 1997) Land measurement
database for 24 surface types in
visible and infrared (NPOESS, Net Heat Flux
ATBD, 2001) - regression method
Integration into CRTM will be completed in June,
2005
11
Cloud optical parameter module

• NESDIS/ORA lookup table (Liu et al., 2005)
  mass extinction coefficient, single scattering
  albedo, asymmetric factor and Legendre phase
  coefficients
• IR spherical particles for liquid water and ice
  cloud (Simmer, 1994) non-spherical ice cloud
  (Liou and Yang, 1995 Macke, Mishenko et al.
  Baum et al., 2001).
• MW spherical particles for rain drops and ice
  cloud (Simmer, 1994).
• Integration with CRTM will be completed in
  June

12
Aerosol optical parameter module

• The initial version includes only dust aerosol
  absorption (no scattering) - aerosol optical
  depth profile (NASA GSFC).
• Integration into the pCRTM (current operational
  RTM) is completed integration with CRTM is
  underway.

13
RT Solution Module

• Four RT solvers being integrated into CRTM
• Solve RT equations for a plane-parallel,
  multiple-layer atmosphere

14
RT Solution Module

• UW Successive Order of Interaction (SOI)
• Truncated doubling technique to compute layer
  transmission, reflection and source functions
  SOS (successive orders of scatterings) to
  integrate emission and scattering events from
  surface to the top of atmosphere (Heidinger et
  al., 2005), IR and MW.
• Forward, tangent-linear and adjoint models.
• The three models have been preliminarily
  integrated with the CRTM framework.

15
RT Solution Module

• NOAA/ETL Discrete-ordinate tangent linear
  radiative transfer model (DOTLRT)
• Matrix operator method to compute layer
  transmission, reflection and source function,
  adding method to combine layers and surface
  (Voronovich et al., 2004), IR and MW.
• Forward and Jacobian models and HG phase function
  lookup table
• Codes were received in February with the DOTLRT
  integrated with an earlier version of the CRTM
  framework (forward interface only). Now ETL is
  revising the codes.

16
RT Solutions (cont.)

• UCLA vector d-4 stream model
• Delta-4 stream algorithm to compute layer
  transmission, reflection and source function
  analytically adding method to combine layers and
  surface (Liou et al., 2005), IR and MW.
• Forward and Jacobian models.
• Forward model is being integrated into CRTM.

17
RT Solutions (cont.)

• NESDIS/ORA Vector DIScrete-Ordinate Radiative
  Transfer (VDISORT)
• Solve for full polarimetric vector, multiple
  stream radiative transfer equation with
  polarization from surface and atmosphere as well
  as their interaction (Weng and Liu, 2003), VIS,
  IR and MW.
• Forward and Jacobian models.
• Forward model integration will be completed in
  June
• Will be used as a benchmark and research tool

18
Plans

• By the end of June, 2005, a beta version CRTM
  will be completed with the following components
• Gaseous absorption modules OPTRAN and OSS if
  completed
• Cloud optical parameter databases ORA and ETL
  lookup tables
• Surface emissivity and reflectivity module with
  LandEM, MW SeaIce/Snow emissivity model, MW Ocean
  emissivity model, IRSSE, and IR land emissivity
  database.
• RT solution modules VDISORT and the following
  modules or programs if completed UW SOI, ETL RT
  Solver and UCLA Vector Delta-4 Stream.

19
Plans (cont.) CRTM test and assessment

• Before passing the CRTMs to the data assimilation
  system for impact evaluation, we will work with
  the community to test and assess the CRTMs for
• (1) software reliability, stability
  and maintainability
• (2) model accuracy
• (3) computation efficiency
• (4) memory use
• Note that we assume the developers
  will fix software bugs and any other deficiencies
  in their codes.
• To test the software and models, we will soon
  provide a set of model inputs including surface
  data for ocean, land, snow, and ice, and profiles
  of temperature, water vapor, ozone, water, ice
  and aerosol parameters.
• We will also provide theoretical results for
  comparisons. Data may be created by LBLRTM and
  VDISORT, or other models such as Doubling-Adding
  method, Monte Carlo methods.
• Sensors AIRS, AMSU, HIRS, and WINDSAT

20
Plans (cont.)

• Testing of the beta version CRTM will be
  completed at the end of September and the tested
  code will be provided to JCSDA.
• Continue to work with the community to integrate
  RT components.
• Conduct comparisons between CRTM calculations and
  observations (CloudSat CALIPSO, ARM, etc.)

21
Issues

• Layer to level profile conversion
• OPTRAN vs. OSS

22
Layer to level profile conversion

• The NWP system produces layer temperature
  profiles, but some RT components require level
  temperature profiles
• Possible solutions
• (1) Assuming Tlayer(i) 0.5(Tlevel(i-1)
  Tlevel(i)), with known Ts and
• Tlayer(i), i1, n, solve the
  equation for Tlevel(i), i0, n
• (2) Predict Tlevel(i), i0, n from Ts and
  Tlayer(i), i1, n using regression
• technique y Ax
• (3) Interpolation

23
Examples of layer to level temperature conversion
The difference between the original level
profile and that retrieved from the layer
profile by solving the equations. 0.5 k error is
added to the surface air temperature.
The difference between the original level
profile and that by interpolating the layer
profile on the level grids. 0.5 k error is added
to the surface air temperature.
Original level profile A layer profile is
constructed from it T_lay(i)
0.5(T_lev(i)T_lev(i1))
24
Comparison between OPTRAN and OSS

• Yong Han, Larry McMillin and Xiaozhen Xiong
• NOAA/NESDIS/ORA
• Jean-Luc Moncet, Gennadi Uymin and Sid Boukabara
• AER, Inc

25
Data sets for the comparisons

• UMBC 101 level 48 profile set
• ECMWF 101 level 52 profile set
• For each set the following data are prepared
• LBLRTM SRF-averaged gaseous transmittances for
  training OPTRAN
• LBLRTM Monochromatic radiances for training OSS
• Ground-truth channel radiances obtained by
  convolving LBLRTM monochromatic radiances with
  the SRFs
• Settings for the independent data set
• Specular surface is assumed IR emissivity
  0.98 MW emissivity 0.6
• Surface pressures are varied among different
  profiles
• Data are prepared (by AER, Inc) for the following
  sensors
• AIRS_aqua, HIRS3_n17, AMSU_n17, SSMIS_f16
• But results shown here only for AIRS, HIRS, AMSU
  and SSMIS

26
Problem in choosing a common training data set

• Initially we want to train and test OPTRAN
  and OSS with the same data sets, but
  unfortunately OPTRAN and OSS are sensitive to
  different issues and therefore have different
  requirements for the training data. OPTRAN is
  better trained with the UMBC set and OSS is
  better trained with five perturbations of the
  ECMWF set.

27
OPTRAN vs. OSS at AMSU channels
OSS Trained with ECMWF set Tested with UMBC set
OPTRAN Trained with ECMWF set Tested with UMBC set
RMS difference
Mean difference
28
OPTRAN-V7 vs. OSS at AIRS channels
OSS Trained with ECMWF set Tested with UMBC set
OPTRAN Trained with UMBC set Tested with ECMWF set
29
Water vapor Jacobians at strong water vapor
channels
30
Water vapor Jacobians at weak water vapor channels
31
Computation Memory Efficiency
Time needed to process 48 profiles with 7
observation angles
Memory resource required (Megabytes)
32
Summary

• Radiance accuracy
• Trained with the ECMWF data set (for a nominal
  accuracy 0.05K) and tested with the UMBC set,
  OSS has an overall accuracy better than 0.05 K
  trained with the UMBC data set and tested with
  the ECMWF data set, OPTRAN has an overall
  accuracy better than 0.1 K
• A good OSS feature is that its radiance accuracy
  can always be improved by increasing the number
  of nodes. However, there is a trade-off between
  the accuracy and the computation and memory
  efficiencies.
• Jacobian accuracy
• Both OPTRAN and OSS provide accurate temperature
  Jacobians and Jacobians for strong absorbers
• The OSS Jacobian model may perform poorly for
  weak absorbers due to the fact that OSS is
  trained in radiance space and the weak absorbers
  are weighted low under the training thresholds
  OPTRAN can provide reasonable Jacobians for weak
  absorbers because OPTRAN is trained in
  transmittance space and errors for each gaseous
  components are minimized.
• Computation efficiency
• OSS is significantly faster than OPTRAN
• Memory requirement
• The amount of memory taken by OSS depends not
  only on the number of channels, but also on the
  degree of node overlap. For the sensors
  considered here, OSS takes significantly more
  memory than OPTRAN.
• Compact OPTRAN is superior in memory use, taking
  only a small fraction of the amount of memory
  required by OSS and OPTRAN-V7.