Evidence that the nature of the boundary

1
LAND-ATMOSPHERE INTERACTION IS THERE ANY
OBSERVATIONAL EVIDENCE?
Findell and Eltahir (1997) provide evidence that
soil moisture variations in Illinois affect
precipitation, though the evidence is disputed
by Salvucci et al. (2002).
Evidence that the nature of the boundary layer
over land is influenced by variations in soil
moisture include the analysis of Betts and Ball
(1995)
dry soil
wet
wet soil
dry
dry
dry
wet
wet
2
Calculate Lag-2 autocorrelationbetween
precipitation pentads
Impacts on precipitation are much more difficult
to identify. Problem The search for evidence of
feedback in nature is limited by scant soil
moisture and evaporation data we have no
observational evidence of feedback on
precipitation at the large scale.

• Question Can we uncover evidence of feedback at
  the large scale in the observational
  precipitation record?

3
Observational data set
Unified Precipitation Database, put together by
Higgins et al.
Daily, ¼ o over the U.S. for 1948-1997
Based on 12000 stations/day (on average)
Assembled from NCD Coop. RFC daily NCDC
hourly, accumulated to daily
Aggregation Data aggregated to pentads (5 day
totals) at 2o x 2.5o.
AGCM strategy for interpreting the observations
1. Identify a feature of interest in the
autocorrelation field (or other field). 2. See
if the AGCM reproduces this behavior. 3. If so,
determine what causes the behavior in the
AGCM. 4. Infer that the same mechanisms apply in
nature.
something of a leap of faith
4
(No Transcript)
5
The observations show a pattern of
autocorrelation that is similar in location and
timing, though not in magnitude, to that produced
by the GCM.
Possible reasons 1. Statistical fluke 2. The
pattern is a reflection of something unrelated to
land-atmosphere feedback, such as monsoon
dynamics, long-term precipitation trends, or SST
variability. 3. The pattern does reflect
land-atmosphere feedback. Note if 3 is
correct, then an analysis of what controls
feedback in the GCM could shed further light on
the observations.
What might be going on? In the west high
evaporation sensitivity yields low soil moisture
memory, and low evaporation yields low impact on
rainfall. In the east consider the
evaporation-versus-soil moisture curve
E
Where things are wet, evaporation is not
sensitive to soil moisture.
W
6
Can we explainwhat controls ac(P) in the GCM?
GCM
obs
correlates with
Pn
Pn2
means that
Pn
Pn2
Breaks down in western US
correlates with
correlates with
En2
correlates with
wn
wn2
Breaks down in eastern US
correlates with
Breaks down in western US
7
Another study Evidence of Feedback in
Observational PDFs
Dataset GPCP monthly precipitation,
1979-2000. Approach Rank precipitation for a
given month into pentiles determine conditional
PDFs of rainfall in the following month for each
pentile. Standardize and assume ergodicity to
generate the PDFs.
Does July rainfall for these years tend to be
higher than normal?
years with highest June rainfall
June rainfall 4th level
June rainfall 3rd level
June rainfall 2nd level
Does July rainfall for these years tend to be
lower than normal?
years with lowest June rainfall
8
(No Transcript)
9
(No Transcript)
10
The AGCM reproduces this observed behavior
11
but only when land-atmosphere feedback in the
model is enabled
12
Note that the broadness of the PDFs implies that
while feedback exists, the prediction skill
associated with the feedback may be quite limited.
13
More studies...
Some AGCM studies examine the impact of
perfectly forecasted soil moisture on the
simulation of observed extreme events. Examples
Hong and Kalnay (Nature, 408, 842-844,
2000) studied the impact of dry soil moisture
conditions on the maintenance of the 1998
Oklahoma-Texas drought.
Schubert et al. (see fig. 1 of Entekhabi et
al., BAMS, 80, 2043-2058, 1999) demonstrated
that their AGCM could only capture the 1988
Midwest drought and the 1993 Midwest flood if
soil moistures were maintained dry and wet,
respectively.
14
Key test Impact of land initialization on
forecast skill
Other studies have examined the impact of
realistic soil moisture initial conditions on
the evolution of subsequent model precipitation.
Douville and Chauvin, Clim. Dyn., 16, 719-736,
2000.
Studies include Viterbo and Betts, JGR, 104,
19361-19366, 1999. Also
Fennessy and Shukla, J. Climate, 12, 3167-3180,
1999.
15
Detailed description of another recent study of
this type (Koster and Suarez, J. Hydromet.,
2003)
POOR MANS LDAS A study of the impacts of soil
moisture initialization on seasonal forecasts
At every time step in a GCM simulation, the land
surface model is forced with observed
precipitation rather than GCM-generated
precipitation. The observed global daily
precipitation data comes from GPCP and covers
the period 1997-2001 at a resolution of 1o X 1o
(George Huffman, pers. Comm.) The daily
precipitation is applied evenly over the day.
16
(No Transcript)
17
Note for the soil moisture initialization runs,
some scaling is required to ensure an initial
condition consistent with the AGCM
Essentially, a dry condition for the GPCP forcing
run
is converted to an equivalently dry condition
for the AGCM forecast simulation.
18
Key finding from this study soil moisture
initialization has an impact on forecasted
precipitation only when three conditions are
satisfied 1. Strong year-to-year variability in
initial soil moisture. 2. Strong sensitivity of
evaporation to soil moisture (slope of
evaporative-fraction-versus-soil-moisture
relationship). 3. Strong sensitivity of
precipitation to evaporation (convective
fraction).
19
On average, there is a hint of improvement associa
ted with land moisture initialization
20
Illustration of point 6 The ensemble mean is
off, but some of the ensemble members do give
a reasonable forecast
21
A more statistically complete experiment was
tried next....
Approach
GLDAS project (NASA/GSFC) using Berg et al.
(2003) data
Wind speed, humidity, air temperature, etc. from
reanalysis
Observed precipitation
Observed radiation
Initial conditions for subseasonal forecasts
Mosaic LSM
The resulting initial conditions (1) Reflect
observed antecedent atmospheric forcing, and (2)
Are consistent with the land surface model used
in the AGCM.
22
1-Month Forecasts Performed
Atmosphere not initialized. Land
initializatized on
May 1 June 1 July 1 Aug. 1
Sept. 1
1979 1980 1981 1992 1993
75 separate 1-month forecasts, each of which can
be evaluated against observations.
(Note each forecast is an average over 9
ensemble members.)
23
We compare all results to a parallel set of
forecasts that do not utilize land
initialization the AMIP forecasts. The AMIP
forecasts do not rely on atmospheric
initialization, either. In essence, the AMIP
forecasts derive skill only from the
specification of SST.
Before we evaluate the forecasts, we ask a
critical question what is the maximum
predictability possible in this forecasting
system? To answer this, we perform an idealized
analysis
STEP 1
For each of the 75 forecasted months, assume that
the first ensemble member represents nature.
STEP 2
For each of these months, assume that the
remaining 8 ensemble members represent the
forecast.
STEP 3
Determine the degree to which the forecast
agrees with the assumed nature.
STEP 4
Repeat 8 times, each ensemble member in turn
taken as nature. Average the resulting skill
diagnostics.
24
Regress forecast against observations to
retrieve r2, our measure of forecast skill.
25
The idealized analysis effectively determines the
degree to which atmospheric chaos foils the
forecast, under the assumptions of perfect
initialization, perfect validation data, and
perfect model physics. In other words, it
provides an estimate of maximum possible
predictability.
26
Where we look for skill is also limited by
quality of observations
27
Areas with adequate idealized predictability and
adequate rain gauge density
Precipitation Forecast Areas
Breadth of areas that can be tested will increase
with future improvements in data collection and
analysis.
Temperature Forecast Areas
28
FORECAST EVALUATION PRECIPITATION
Without initialization
With initialization
Differences
Idealized differences
29
FORECAST EVALUATION TEMPERATURE
Without initialization
With initialization
Differences
Idealized differences
30
(No Transcript)
31
What happens when the atmosphere is initialized
(via reanalysis) in addition to the land
variables? Supplemental 9-member ensemble
forecasts, for June only (1979-1993) 1.
Initialize atmosphere and land 2. Initialize
atmosphere only
Warning Statistics are based on only 15 data
pairs!
June r2 values, averaged over area of focus
AMIP runs SSTs only
SSTs land initialization atmosphere initializ
ation
SSTs atmosphere initialization
GLDAS runs SSTs land initialization
32
June r2 values, averaged over area of focus
AMIP runs SSTs only
SSTs land initialization atmosphere initializ
ation
SSTs atmosphere initialization
GLDAS runs SSTs land initialization
33
Outlook
Presumably, skill associated with land
initialization can only increase with --
improvements in model physics -- improved data
for initialization satellite
sensors (HYDROS, GPM, ) ground networks data
assimilation -- improved data for validation In
other words, weve demonstrated only a minimum
skill associated with land initialization.
Current increase in skill
Idealized potential increase in skill
We have a lot of untapped potential!
34
DATA ASSIMILATION THE OPTIMAL MERGING OF
OBSERVATIONS AND MODEL RESULTS
strengths
weaknesses
Key motivation observations and model results
have their own strengths and weaknesses. By
combining them optimally, we get the best of both
worlds.
Measures of real-world states. What were after
in the first place.
Inadequate coverage, significant measurement
error.
obs
Based on trustworthy physics. Complete space/
time coverage, including unmeasurable states.
Results subject to myriad inadequacies of model
parameterizations
model
Decide how much you believe model result
(estimate model error)
Integrate model forward in time
Combined model/ observational state at time t
Combined model/ observational state at time t1
Model state at time t1
Decide how much you believe obser-vation
(estimate observational error)
Observation at time t1
The actual mathematics involved here can be very
complicated...
35
Retrospective period
SMMR data
assimilation
(scaling)
Climatology / interannual variability of land
surface states
Corrected ECMWF reanalysis
LSM
forcing
Real time forecasts
scaling
AMSR data
assimilation
(scaling)
Initial conditions for forecasts
GLDAS
LSM
forcing