Geoscience Reference
In-Depth Information
the physical modelling of complex processes in snow, and remote sensing products are
global, but limited by signal saturation and do not provide a unique SWE solution on
inversion due to their high sensitivity to other snow properties.
Data assimilation techniques that use microwave information to update a forecast from
other sources have been suggested to improve snow mass estimation. ESA's Globsnow
uses modern assimilation techniques to bring together ground observations and remote
sensing products and has shown that assimilating microwave measurements does improve
SWE estimates. Globsnow isolates and accounts for the snow microstructure's contribution
through a grain size parameter, which is obtained by fitting ground measurements to
satellite retrievals while assuming a single homogeneous snow layer. The brightness
temperature observable chosen by Globsnow is the difference between brightness tem-
peratures at 19 and 37 GHz vertically polarised microwaves, DT
B
;
V
.
The Globsnow grain size estimate is reliant on point measurements of snow, which may
vary greatly over relatively small areas, and it is suggested that physically based snow
models could provide an alternate source of information to improve the inversion of the
passive microwave signal.
Snow forms in layers and its stratigraphy can be complex, though physical models are
capable of reproducing this layering. Lemmetyinen et al. (
2010
) and Durand et al. (
2011
)
showed that for simulation of brightness temperatures over small areas of snow, this
complexity can be an important contribution to the signal. Globsnow ignores this com-
plexity in determining the error covariance for weighting the observational increment in
the update step, and this might lead to suboptimal updates.
The HUT radiative transfer model used in Globsnow was able to simulate satellite-
observed DT
B
;
V
with an RMSE of 8 K, down from the 17 K RMSE associated with
estimates made using the Chang algorithm typical of stand-alone microwave SWE pro-
ducts. The HUT RMSE was 6 K excluding one site where it was believed that snowpits
were biased towards thin snow.
After confirming that the HUT radiative transfer model used in Globsnow was able to
simulate satellite-observed scene brightness temperatures at NASA's CLPX, the HUT-
simulated DT
B
;
V
values for CLPX snowpits resampled to different layering structures were
compared. Simulated DT
B
;
V
for snow with the maximum possible level of stratigraphic
detail based on the 10-cm resolution of CLPX density and temperature measurements was
taken as truth, and deviations from this were treated as due to errors introduced by sim-
plification of the stratigraphy to fewer layers.
Removing layering detail leads to a bias in the simulated DT
B
;
V
, likely due to the
removal of reflection effects at layer boundaries and possibly due to nonlinearities in the
DT
B
;
V
response to snow grain size and density. Globsnow can freely vary the grain size
to account for this, but this is likely to have second-order effects on the assimilation
scheme.
Simulated DT
B
;
V
values for the same snowpit at different levels of layering detail were
found to vary, with the standard deviation increasing approximately linearly with snow
depth. For snow of depth 100 cm (172 mm SWE at the CLPX sites), the standard deviation
in simulated DT
B
;
V
values for a single-layer model versus the N-layer model was estimated
at 4.8 K, equivalent to approximately 13 mm SWE (7 % of total). Using two snow layers
reduced the DT
B
;
V
error to 2.1 K (5.6 mm SWE, 3 % of total).
Globsnow reports RMSE values of 40 mm for SWE \150 mm using a single-layer
version of HUT, and the values found here suggest that layering could be a notable
component of that RMSE.
