NASA GEOS-3/TRMM Re-Analysis:
Capturing Observed Rainfall Variability in Global Analysis
Arthur Hou
NASA Goddard Space Flight Center
2nd IPWG Workshop, Naval Research Laboratory
Monterey, CA, 25-28 October 2004
Hou/JTST2000 - 1
Scope of talk

Precipitation products from most operational NWP systems are
forecasts rather than analyses of precipitation based on rainfall
observations and model information.

NASA has been exploring a different approach to precipitation
assimilation that uses rainfall observations to directly estimate and
correct errors in the model rain within a 6h assimilation cycle.

A brief description of the variational continuous assimilation
(VCA) scheme for precipitation assimilation.

Results from the NASA GEOS-3/TRMM reanalysis (Nov. 1997Dec. 2002):
- An atmospheric analysis dynamically consistent with a QPE based
on TMI and SSM/I rain rates.
GEOS-3 = Goddard Earth Observing System – Version 3
Arthur Hou, IPWG Workshop, 25-28 October 2004 - 2
Tropical ENSO rainfall variability: Observation vs analyses
GPCP
TMI Monthly-Mean SST
January: 1998 Minus 1999
January 1998
January 1999
NCEP
January: 1998 Minus 1999


ERA40
January 1998 Minus 1999

Large discrepancies (for the same SST input)
Tropical rainfall analyses are model-dependent
and vary with parameterized model physics
Present-generation convective schemes are less
than perfect - systematic model errors
1 mm/day ~ 30 W/m2
mm/day
Arthur Hou, IPWG Workshop, 25-28 October 2004 - 3
Sensitivity of tropical analysis to precipitation process
Variance of Hadley circulation streamfunction
ECMWF Reanalysis (80-93)




NCEP/NCAR Reanalysis
September 1982: Diabatic Nonlinear
Normal Initialization (DNNMI)
implemented at ECMWF
September 1984: DNNMI introduced
at NMC
May 1985: Shallow convection (SC)
implemented at ECMWF
May 1986: SC implemented at NMC
GEOS-1 Reanalysis
Time series of 15-d mean of tropical [v] at 200 hPa
Trenberth & Olson, 1988
Arthur Hou, IPWG Workshop, 25-28 October 2004 - 4
Key issues in precipitation assimilation

Conventional data assimilation algorithms are based on the
assumption that the underlying observation and model error
statistics are random, unbiased, stationary, and normally distributed.

But model clouds and precipitation are derived from parameterized
moist physics, which can have large systematic errors. Unless these
(largely unknown) systematic model errors are accounted for in the
assimilation procedure, one will always make sub-optimal use of
these data.

A basic problem is that the observation operator for precipitation is
not as accurate as those for conventional data or observables in
clear-sky regions.
Arthur Hou, IPWG Workshop, 25-28 October 2004 - 5
What is an observation operator?
It relates an observable to model state variables (u,v,T,q, etc.)
(u,v,T,q)model grid
random error
Observation
operator
“Perfect model”
(T,q,u,v)grid
Precipitation
Precipitation
observation operator
observation operator
with correction, d
(u,v,T,q)
Observations
at
observation
in clear-sky
locations
regions
Systematic error
Cloud,
Precipitation
Developing procedures to make online estimation and correction of biases in
the observation operator to make more effective use of precipitation data
Arthur Hou, IPWG Workshop, 25-28 October 2004 - 6
Variational continuous assimilation (VCA) of surface rain
•
The strategy is to relax the perfect model assumption - i.e., using the
forecast model as a weak constraint.
•
Assimilation of 6h surface rain accumulation using 6h-mean moisture
tendency correction as the control variable, and applying the correction
continuously over a 6h analysis window to ensure dynamical consistency.
Minimizing the cost function:
A 1+1D observation operator (H) based
on a 6h time-integration of a column
model of moist physics with large-scale
forcing prescribed from “first guess”
•
J(x) = (x)T P-1 (x) + ( yo – H(x))T R -1 ( yo – H(x))





model tendency correction: x
logarithm of observed rain rate: yo
logarithm of model rain estimate: H(x)
error covariance of prior estimate: P
logarithm of relative observation error variance: R
The scheme estimates and corrects for biases in model’s moisture tendency
every 6h to minimize discrepancies in 6h rain between the model and
observations.
Arthur Hou, IPWG Workshop, 25-28 October 2004 - 7
Impact of VCA rainfall assimilation on GEOS-3 analysis
Replicating
Propagationobserved
and intensity
propagation
of tropical
and intensity of
tropical
rainfall rainfall
systemssystems
are difficult
and intraseasonal
to capture oscillation
GPCP
NCEP GDAS
ERA-40
GEOS/TRMM
mm/d
Rain error reduction (30N-30S, ocean)
GEOS = Goddard Earth Observing System
MJO in precipitation over tropical oceans (10N-10S) 2001
Arthur Hou, IPWG Workshop, 25-28 October 2004 - 8
Improved temporal and spatial variability
Avg. Precipitation (120-150E, 4S-4N)
(Morlet analysis)
Enhanced frequency-time coherence
between GPCP and GEOS-3 analysis
An atmospheric analysis dynamically consistent
with observed rainfall variability
Arthur Hou, IPWG Workshop, 25-28 October 2004 - 9
Improved cloud radiative forcing verified against CERES
Variational continuous
rainfall assimilation
improves key climate
parameters such as clouds
and TOA radiation in the
GEOS analysis


94% reduction in bias
51% reduction in error
standard deviation
January 1998
Arthur Hou, IPWG Workshop, 25-28 October 2004 - 10
Impact on wind and humidity analyses
Improved latent heating patterns and large-scale motion fields leading to improved
upper-tropospheric humidity (verified against TOVS brightness temperature)
GEOS(TMI+SSM/I PCP+TPW) minus GEOS(CONTROL)
Verification: HIRS2 Channel 12 Brightness Temperature
Surface rain
& Horizontal
div. wind
at 200 hPa
Omega
velocity
at 500 hPa
Specific
humidity
at 400 hPa
January 1998
GEOS control has a moist/cold bias
relative to HIRS2 channel 12 (top)
Rainfall assimilation leads to a drier
upper-troposphere & reduces the
err.std.dev by 11%
Arthur Hou, IPWG Workshop, 25-28 October 2004 - 11
Impact on hurricane track and precipitation forecasts
Bonnie
Floyd
5-day track forecast from 12UTC 8/20/98
5-day track forecast from 00UTC 9/11/99
Improved initial
storm position
5-day rain forecast
5-day rain forecast
Blue: No rainfall data in IC
Red: With rainfall data in IC
Green: NOAA “best track”
Hou et al. 2004:
MWR, August issue.
Arthur Hou, IPWG Workshop, 25-28 October 2004 - 12
Assimilation of TMI, SSM/I & AMSR-E rain
Precipitation July 2002
mm/d
OLR July
OSR
July 2002
2002
W/m2
Arthur Hou, IPWG Workshop, 25-28 October 2004 - 13
Summary

Optimal use of precipitation information in global data assimilation
poses a special challenge because parameterized physics can have large
systematic errors, which must be accounted for in the assimilation
procedure.
– One effective strategy is to assimilate rainfall data using the forecast
model as a weak constraint
– Exploring advanced techniques such as ensemble DA, which could
provide a unified framework for addressing both initial-condition errors
and model errors

The GEOS-3/TRMM reanalysis provides an atmospheric analysis
dynamically consistent with the observed tropical rainfall variability:
– Improved climate parameters including TOA radiation, uppertropospheric humidity, and cloud-radiative forcing
– Improved short-range forecasts
Arthur Hou, IPWG Workshop, 25-28 October 2004 - 14