Geoscience Reference
In-Depth Information
The calibrated parameters are shown in Fig. 6 . Results suggest, for example, that the
roughness parameter (h) is too low in Lit1 and too high in Lit3. The calibrated vegetation
opacity (s) values distinguish clearly between high and low vegetation. The calibrated
scattering albedo (x) is increased over low vegetation, which reduces the vegetation effect
in the simulated Tb. In summary, the climatological calibration generates plausible
parameter values that are consistent with the underlying land modeling system.
3.3.2 Predicting Microwave Brightness Temperatures over Snow
As demonstrated in the previous section, the Catchment model (as do similar global land
surface models) supports the application of a physically based microwave RTM for warm-
season processes. However, the snow model components in global land surface models,
including that in the Catchment model, are usually too simplistic to support physically
based RTM modeling in the presence of snow. Specifically, global snow models lack
reliable estimates of snow microphysical properties (such as grain size, ice layers, and
depth hoar) which would be needed for physically based forward modeling of the
microwave brightness temperatures. Forman et al. ( 2013 ) therefore constructed an
empirical forward RTM for snow-covered land surfaces based on an Artificial Neural
Network (ANN).
The Catchment model state variables used as input to the ANN include the density and
temperature of the snowpack at multiple depths, the temperature of the underlying soil, the
overlying air, and the vegetative canopy, and the total amount of water equivalent within
the snowpack. In addition, a cumulative temperature gradient index (TGI) is used as a
proxy for snow grain size evolution in the presence of a vapor pressure gradient. Using the
above inputs, the ANN is trained and (independently) validated using 10.7, 18.7, and
36.5 GHz microwave brightness temperatures at H- and V-polarization from AMSR-E.
The independent validation is accomplished as follows: From the 9-year AMSR-E data
record, each single year is withheld in turn from the ANN training, and skill metrics for the
resulting ANN predictions are computed only against the AMSR-E data that have been
withheld from the ANN training.
Figure 8 demonstrates the performance of the ANN predictions relative to AMSR-E
measurements that were not used during training. The figure illustrates the overall ability
of the ANN to predict Tbs for the 10 GHz V-polarized channel. The ANN predictions are
essentially unbiased (relative to the AMSR-E measurements) across the 9-year period
(Fig. 8 a). The RMSE is typically less than 5 K (Fig. 8 b). In addition, the ANN demon-
strates skill in predicting interannual variability, with anomaly R values well above 0.5
over large parts of North America (Fig. 8 c). Relatively low skill can be seen in areas along
the southern periphery, where the snowpack is relatively thin and ephemeral, as well as in
areas north of the boreal forest, where sub-grid scale lake ice (which is not modeled in the
land surface model) is common. In short, Fig. 8 suggests considerable skill by the ANN at
predicting interannual variability in 10 GHz V-polarized Tbs across North America with
negligible bias and a reasonable RMSE. The RMSE is somewhat higher but still reasonable
(less than 10 K) for the higher frequencies and for H-polarization Tb (see Figures 4-6 of
Forman et al. 2013 ).
Forman et al. ( 2013 ) also assessed the potential for using the ANN as a forward
observation operator in radiance-based snow assimilation. For this demonstration, the
observations are considered to be in the form of spectral differences in V-polarization
brightness temperatures, DTb : Tb V (18 GHz) - Tb V (36 GHz). Since DTb typically
increases with increasing SWE, this spectral difference is commonly used to estimate SWE
Search WWH ::




Custom Search