Geoscience Reference
In-Depth Information
Metrics used for skill assessment include the bias, root mean square error (RMSE), and
time series correlation coefficient (R). When specified, anomalies were computed by
removing a seasonally varying climatology from the data before computing the metrics.
3 Results
3.1 Assimilation of Sparse and Coarse-Scale Observations
Snow is an important component of the land system because of its strong impact on the
land surface water and energy balance, weather, climate, and water resources (Barnett et al.
2005
). However, land surface models often represent snow processes poorly. Satellite
observations of SWE can be retrieved from passive microwave sensors, but they are only
available at relatively coarse resolution. Moreover, SWE retrievals, like most satellite
observations, do not provide complete spatial and continuous temporal coverage due to
orbit or sensor limitations. The challenge is therefore to design an assimilation system that
can use coarse-scale satellite observations to provide enhanced model estimates at the finer
scales of interest (horizontal downscaling) and that can also propagate the information to
intermittently unobserved areas.
Using AMSR-E SWE retrievals and MODIS SCF observations, De Lannoy et al. (
2010
,
2012
) developed a data assimilation and downscaling technique for estimating fine-scale
(1 km) snow fields using coarse-scale (25 km) SWE retrievals and fine-scale (500 m) SCF
retrievals for a domain in Northern Colorado, USA. In their study, the authors used the LIS
version of the GEOS-5 EnKF together with the Noah land surface model (Ek et al.
2003
)
(rather than the GEOS-5 LDAS and the Catchment model used elsewhere in this paper).
The Noah model simulates a single snow layer with two prognostic variables for SWE and
snow depth. The default LIS soil, vegetation, and general parameter tables for Noah were
used, including a Noah-specific maximum snow albedo.
Figure
1
shows schematically how the coarse-scale SWE retrievals are used. The fine-
scale model grid is represented by the dashed lines in the figure. The coarse-scale grid of
the SWE observations is represented by the solid lines and light/dark gray shading, and the
center points of individual SWE retrievals are marked with crosses. Let us now consider
the analysis update of the fine-scale model grid cell indicated by the solid black square.
First, it is important to emphasize that the coarse-scale SWE retrievals are not compared
directly to the SWE estimate at the fine-scale model grid cell. Rather, the model SWE is
aggregated to the coarse grid of the retrievals, that is, the fine-scale model forecast is
mapped into the coarse-scale observation space. This aggregation is part of the observation
operator that maps the model states to the observations. Observation-minus-model-forecast
residuals (or innovations) are then computed at the coarse scale of the observation space.
The Kalman gain matrix transforms the (observation-space) innovations into the (model-
space) increments. It is computed from error correlations between the model states at the
fine scale and the model-predicted measurements at the coarse scale. Finally, the incre-
ments are added to the (fine-scale) model forecast in the analysis update. See De Lannoy
et al. (
2010
) for a discussion based on equations.
Second, multiple coarse-scale SWE retrievals in the vicinity of the fine-scale model grid
cell in question are used for the analysis update. Specifically, the update uses the three
coarse-scale SWE retrievals marked by black crosses that are within a given radius
(indicated by the white semi-circle) around the fine-scale model grid cell in question
(Fig.
1
). Note that this model grid cell would be updated even if the SWE retrieval directly
