A Land Surface Model Hind Cast for the Arctic Terrestrial Drainage Basin

1
A Land Surface Model Hind Cast for the Arctic
Terrestrial Drainage Basin Theodore J. Bohn1, And
rew G. Slater2, James McCreight2, Dennis P.
Lettenmaier1, and Mark C. Serreze2
1Department of Civil and Environmental
Engineering, Box 352700, University of
Washington, Seattle, WA 98195
2Cooperative Institute for Research in
Environmental Sciences, 216 UCB, University of
Colorado, Boulder, CO 80309-0216
Arctic CHAMP Freshwater Initiative Conference (J
une 1-3 2005)
4
Comparison of ERA-40-based and NCEP-based Forcings
ABSTRACT River runoff from the Arctic terrestrial
drainage system is thought to exert a
significant influence over global climate,
contributing to the global thermohaline
circulation via its effects on salinity, sea ice,
and surface freshening in the North Atlantic.
Changes in these freshwater fluxes, as well as
other components of the Arctic terrestrial
hydrologic cycle such as snow cover and albedo,
have the potential to amplify the Arctics
response to global climate change. However, the
extent to which the Arctic terrestrial
hydrological cycle is changing or may contribute
to change through feedback processes is still not
well understood, in part due to the sparseness of
observations of such variables as stream flow,
snow water equivalent, and energy fluxes. The
objective of this project is to assemble the best
possible time series (covering a 20 year period)
of these and other prognostic variables for the
Arctic terrestrial drainage basin. While these
variables can be estimated with a single land
surface model (LSM), the predictions are often
subject to biases and errors in the input
atmospheric forcings and limited by the accuracy
of the model physics. To reduce these errors, we
have followed a two-pronged approach on one
hand, we are assembling an optimum set of
atmospheric forcings by blending the output of
the ERA-40 reanalysis project with gridded
in-situ and satellite observations on the other
hand, we have created an ensemble of five LSMs
VIC1, CLM2, ECMWF3, NOAH4 and CHASM5, all of
which have been used previously to simulate
Arctic hydrology under the Project for
Intercomparison of Land-surface Parameterization
Schemes (PILPS) Experiment 2e. Model predictions
of land surface state variables (snow water
content, soil moisture, permafrost active layer
depth) and fluxes (latent, sensible, and ground
heat fluxes runoff) are averaged both across the
ensemble and over multiple runs, using our
optimum atmospheric forcing data with and without
added random perturbations. Here we demonstrate
the results from an individual model using our
ERA-40-derived forcings, and evaluate the
performance of the multi-model ensemble averages
in comparison with individual model simulations
of variables including snow water equivalent and
total runoff over the pan-arctic domain (using
forcings derived from Adam and Lettenmaier (2003)
and NCEP reanalysis). 1Variable Infiltration
Capacity macroscale model (Liang et al.,
1994) 4NCEP, OSU, Air Force, and NWS Hydrologic
Research Lab collaborative model
2Community Land Model (NCAR UCAR) 5CHAmeleon
Surface Model 3European Center for Medium-range W
eather Forecasting, land component of Integrated
Forecast System model
Given the influence of forcing data, we are
compiling an ensemble of forcings which includes
assimilated station based data, satellite
radiances and samples from the NCEP reanalysis.
Figure 2 shows examples of differences in mean
monthly incident shortwave and longwave
radiation. The month shown is June, thus
differences will be near a maximum. We have
fields that are based on ERA-40, but include
satellite estimates of cloud cover (Fig 2a). The
suggested bias in the longwave flux is clearly
shown. An alternative set of radiation fields is
based on NCEP reanalysis (Fig 2b) and shows a
different spatial distribution of differences.
A
1
2
Typical LSM Structure
Ensemble Process Flow
The LSMs in our ensemble all share the same basic
structure, consisting of grid cells containing a
multi-layer soil column overlain by one or more
tiles of different land covers, including
vegetation with and without canopy, bare soil,
and in some cases, lakes, wetlands, or glaciers.
Water and energy fluxes are tracked vertically
throughout the column from the atmosphere through
the land cover to the bottom soil layer. The
figure to the right illustrates these features as
implemented in the VIC (Variable Infiltration
Capacity) macroscale land surface model (Liang et
al., 1994).
Process flow in the multi-model ensemble.
Forcing data consist of ALMA variables stored in
NetCDF format files. These are translated into
each models native variables and format. After
the model simulations finish, the results are
translated back into ALMA-standard variables and
stored in NetCDF files. These standardized
results are then analyzed over a training
period and combined to form the ensembles
aggregate result.
B
Figure 1 Simulated latent heat fluxes for the
CHASM model using ERA-40 based forcings.
Figure 1 shows simulated latent heat fluxes for
the CHASM model using ERA-40 based forcings. Two
versions of the model are shown. The 1 Tile
version (upper panel) computes a single energy
balance for the grid box and lumps bare soil and
vegetation together to form an effective
parameter model. In contrast, the 2 Tile version
(mid panel) computes a separate energy balance
for each of the vegetated and bare soil portions.
Notably more evaporation occurs in the
mid-latitude regions with the 1 Tile model.
Overall, the models produce similar results to
those taken directly from ERA-40 (lower panel)
and suggests that the forcings play a major role
in determining the resultant fluxes.
Figure 2 Differences in shortwave and longwave
radiation fields between ERA-40 and (a) AVHRR
cloud-cover-corrected and (b) NCEP-based forcings.
3
Ensemble Results using Adam/NCEP Forcings
CONCLUDING REMARKS While this study is still in i
ts preliminary stages, evidence so far suggests
that An ensemble of land surface models can make
more accurate predictions of hydrological
variables than individual models.
An ensemble trained against one variable (in this
case fractional snow cover) can make plausible
predictions of other variables (e.g. annual
discharge). This may not be true in general,
however, and deserves further investigation.
The outputs of individual models can be sensitive
to choice of forcing dataset. A multi-model
ensemble may be more stable than individual
models in the face of different forcings, due to
the compensatory effects of the model weights.
However, any systematic responses across all
models would remain. Ensembles of land surface mo
dels may be able to help us more accurately
estimate hydrological variables in regions where
there are few observations. Future Work The e
nsembles sensitivity to choice of training
variable remains to be determined.
Implementation of multiple constraints (e.g. snow
cover and stream flow simultaneously).
Testing of weights that vary in time and space.
Investigation of the effects of data assimilation
on ensemble performance. Note See the author
for a list of references.
Using the coefficients derived from the linear
regression against fractional snow cover, we have
combined the models results for annual discharge
in the Lena, Mackenzie, Ob, and Yenisei river
basins, shown at right. For three of the four r
ivers, the ensemble does at least as well or
better than the individual models at predicting
annual discharge. In the case of the Ob, the
ensemble is clearly influenced by extreme results
from two of its members. We have not yet
investigated the cause of this large discrepancy,
but a possible cause is over-prediction of
condensation in at least one model.
It should be noted that the relative performance
of the individual models for prediction of
annual stream flow is not necessarily the same as
for fractional snow cover. For example, our
implementation of CHASM predicts snow cover
relatively well but predicts discharge relatively
poorly. However, the relative performance of th
e individual models seems to vary little from
basin to basin. The main exception is CHASMs
performance in the Mackenzie basin, which agrees
with observations much better than elsewhere.
Whether this is part of a systematic difference
between North America and Eurasia remains to be
seen.
Ensemble results are a linear combination of the
results of the individual LSMs, whose weights are
the coefficients from a multiple linear
regression of observations against the model
predictions. In this example, we trained the
ensemble against weekly MODIS snow cover extent,
running the models with the NCEP-based forcings
of Adam and Lettenmaier (2003).
The rms error between the ensemble results and
observations, as well as the rms errors of the
individual models, are shown below. In general,
the ensemble rms error tends to be less than or
equal to the lowest of the rms errors of the
individual models.
Representative snapshots of each season (labeled
A , B, C, and D) are displayed to the
right, for observations and ensemble predictions
of fractional snow cover, and the error
(predicted minus observed). High rms error tends
to occur in spring and fall at the edges of the
snow-covered region. This is due in part to the
fact that the model results used here are
3-hourly data averaged to weekly intervals, which
tends to blur the boundary between snow-covered
and snow-free regions, while the observations are
weekly snapshots. A more fair comparison would
use sub-sampled model results to mimic the
observations.