ANalysis of Snow Water Equivalent Retrievals (ANSWER)

by M. Sandells, N. Rutter, L. Wake

The Analysis of Snow Water Equivalent Retrievals (ANSWER) project enabled the UK to
contribute to the EE10 ‘Radar Imager for Sensing the Cryosphere (RISC)’ mission concept
proposal as a member of the science team with simulations to support mission feasibility.
RISC is a dual Ku-band (13.5 and 17.2 GHz) active instrument concept at VV and VH
polarizations, designed to provide information on snow water equivalent (SWE), ocean wind
speeds, and sea ice. It has an imaging resolution of 250m with an option for high resolution
(~40m) in mountainous regions, and a range of incidence angles from 23-50o. Two retrieval
methodologies were proposed for SWE: a direct SWE product from backscatter and
brightness temperature, and a SWE product derived from land surface data assimilation.
For the RISC mission concept proposal, the ANSWER project provided: 1) SWE error
budget analysis for uncertainties in snow microstructure parameter estimation, and 2) the
impact of spatial variability in snow properties on SWE retrieval given the proposed RISC
direct SWE retrieval methodology.

ANSWER used data from a field campaign in Trail Valley Creek (TVC), near Inuvik,
NWT, Canada that took place in April 2013. In this experiment, a 50 m trench and eight 5 m
trenches were dug across the tundra site. Stratigraphic layers were imaged across the
trench with near-infrared photographic techniques, and multiple, spatially distributed vertical
profiles were measured focusing on snow microstructure (Specific Surface Area, SSA) and
density. From this, three layer types were identified: fresh snow at the surface (large SSA,
low density); wind slab (mid-range SSA, high density) and depth hoar at the base (small
SSA corresponding to large, poorly bonded crystals with mid-range density). Mean and
standard deviation of SSA and density were calculated for each layer, and relationships
between layer thickness and snowpack depth were determined. These TVC data were used
to quantify SWE retrieval error budget contributions with the Snow Microwave Radiative
Transfer (SMRT) model.

SMRT is a new active/passive microwave scattering model for snow, which has a
modular structure and encompasses numerous previously developed and commonly used
microwave emission models. The modular structure means that the snow microstructure
assumptions can be separated (though must remain compatible with) the electromagnetic
theory and neither are specific to the assumptions used to solve the radiative transfer
equation. SMRT was used to provide forward ‘truth’ simulations for an idealised, synthetic
scene, then applied iteratively to find an optimal match to ‘truth’ backscatter through cost
function minimisation under simplified conditions or with an error applied.

The synthetic scenarios covered three main retrieval assumptions: single (or dual)
layer approximation to a three-layer snowpack, uncertainty in the a priori parameter
estimate used in the single-layer snowpack SWE retrieval, and homogeneity in the retrieval
for a heterogeneous truth. Simplification of a three-layer snowpack to a single layer
snowpack with a priori depth information showed it was possible to obtain an accurate
retrieval of SWE, although in general the SWE was underestimated in part due to the
assumed density in the retrieval of an effective microstructure. Without a priori depth
information, the retrieved depth is highly dependent on the microstructure parameter. For
two-layer retrievals, use of the typical density in the retrieval may also be the reason for low
estimates of the microstructure parameter in either or both layers.

Retrieval accuracy is highly dependent on the microstructure parameter. An
accuracy of 10% or better is required to remain within the RISC error budget limit of 30mm
SWE. This is comparable to measurement accuracy. Spatial variability in microstructure
has little impact in comparison with the a priori error, although SSA smaller than the mean
(equivalent to larger optical radius as these are inversely proportional) may be required to
mitigate spatial variability effects of the microstructure. A priori density error has little impact
on the SWE error budget, although the effect of spatial variability in density (and therefore
SWE) is larger. SWE error budget is maintained with up to a 5K error in a priori physical
snowpack temperature. Spatial variability in temperature and snow depth have no effect on
the retrieved SWE. A large variation in backscatter difference was found for different SMRT
configurations, possibly due to non-equivalence of microstructure models. This warrants
further attention, as does the ability of snowpack models to provide the necessary a priori
data with sufficient accuracy for successful retrievals.

Should RISC be selected by ESA for Phase A studies, the ANSWER methodology
may be expanded to test different forms of the backscatter cost function, investigate the
potential benefits of two-layer retrievals and examine retrieval performance for different
snow climatologies. ANSWER has improved the SMRT model through identification and
resolution of an issue with the radiative transfer solver under specific conditions, and
development and successful testing of SMRT parallelization. The UK capability to
undertake rapid, large-scale sensitivity analyses as would be required for Phase A studies
has been increased through parallelization of SMRT on the High Performance Cluster at
Northumbria University.