Geoscience Reference
In-Depth Information
The importance of correctly specifying random errors and biases is a major conceptual
challenge in the optimization of distributed assimilation systems. Bias mitigation has
become a regular part of most soil moisture data assimilation systems (Reichle and Koster
2004
; Drusch et al.
2005
; Kumar et al.
2012
; Sahoo et al.
2013
), and random error
specifications for soil moisture data assimilation have been studied through adaptive fil-
tering (Crow and van Loon
2006
; Reichle et al.
2008
).
Another idea with potential benefit is multi-sensor assimilation for soil moisture esti-
mation. As an example, Draper et al. (
2012
) showed how both active (ASCAT) and passive
(AMSR-E) microwave retrievals can contribute to a similar improvement in assimilation
results. Combining improved precipitation data with soil moisture retrieval assimilation
(Liu et al.
2011
) and combining discharge (Pauwels and De Lannoy
2006
), temperature or
LAI with soil moisture assimilation are other avenues that have been exploited for
hydrological assimilation.
As already indicated for single-column applications, a major conceptual problem is the
direct assimilation of brightness temperatures (T
b
) or backscatter observations from
satellite missions for soil moisture estimation. Reichle et al. (
2001a
,
b
) presented pio-
neering studies with a 3-D variational scheme to assimilate and disaggregate synthetic or
real brightness temperatures over the SGP97 study area, while Margulis et al. (
2002
) used
an EnKF and Dunne and Entekhabi (
2006
) compared an EnKF with an EnKS for the same
T
b
assimilation problem. Walker et al. (
2002
) also assimilated T
b
directly, but from SMMR
and using an EKF over Australia. Using a variational scheme, and with inclusion of both a
land surface temperature and microwave brightness temperature observation operator,
Barrett and Renzullo (
2009
) showed that both thermal (AVHRR) and microwave (AMSR-
E) satellite observations can provide effective observational constraints on the modelled
profile and on surface soil moisture. There are only a few studies on spatially distributed
backscatter assimilation, but in a recent OSSE using an EnKF, Flores et al. (
2012
) showed
the potential of the L-band radar information expected from the future SMAP mission.
For snow, spatially distributed assimilation studies include snow cover area (or snow
cover fraction) and snow water equivalent (SWE) assimilation. A correct specification of
the snow-covered area is important to represent feedbacks from the land to the atmosphere,
while a good estimate of the actual amount of snow in the snowpack is of crucial
importance for flood, drought and discharge predictions (He et al.
2012
). Snow cover
observations are typically fine-scale visible/near infrared observations that are only
available in cloud-free areas, while SWE measurements are typically more inaccurate
retrievals from T
b
observations at a coarse scale (see Table
1
). It can be expected that
multi-sensor assimilation could help to further snow estimation (De Lannoy et al.
2012
).
Because of its binary nature, snow cover in terms of the presence or absence of snow
cannot be assimilated with filters that rely on continuous variables. Instead, rule-based
algorithms have been proposed (Rodell and Houser
2004
; Zaitchik and Rodell
2009
; Roy
et al.
2010
). However, the snow cover fraction (SCF) is a more continuous variable that has
been assimilated with KF-based algorithms (Clark et al.
2006
; Su et al.
2008
; De Lannoy
et al.
2012
). When assimilating SCF with a Kalman filter, there is a need to relate SCF to
the actual SWE state variable through an observation operator, often defined as a snow
depletion curve. It is also possible to use visible/near infrared snow albedo observations to
update snow parameters such as grain size (Dumont et al.
2012
).
The two dominant conceptual problems with satellite-based SWE assimilation are the
coarse-scale nature and high uncertainty of the measurements. Initial attempts to assimilate
SMMR or AMSR-E SWE retrievals only yielded marginal success (Andreadis and
Lettenmaier
2006
; Dong et al.
2007
), because of retrieval errors due to signal saturation,
