Geology Reference
In-Depth Information
of YI caused by densely packed pancake ice. It can also
be attributed to its relatively high surface salinity that
gives rise to higher permittivity, causing more surface
scattering. The backscatter from the OW is much less
than that from any ice type and is unaffected by surface
wind. These are two advantages of the Ku‐band com-
pared to the C‐band. It is worth noting the similarity
between the distributions of
0.6
0.5
0.4
0.3
0.2
0
. It can be argued
that using uncorrelated parameters as input to ECICE
would be better. Selection of input parameters should
be based on visual examination of the probability distri-
bution functions. Parameters with less correlation and
less overlapping between histogram surface types are
preferred.
The optimal solution in ECICE is applied to the
same set of observations for each pixel using each set
of the CRV (once again, the number is selected by the
user). This results in a number of possible solutions
equal to the number of the selected CRVs. To deter-
mine the final answer from the possible solutions, the
algorithm examines the mode and the median of the
probability distribution of the output ice concentra-
tion vectors from all possible solutions. The median is
selected as being the final answer unless the mode of
the concentration of at least one surface is 100%. In
that case, the median has to be incremented to bring its
value close to 100% concentration for the surface that
has its highest mode at 100% concentration. This will
constitute the final answer of the concentration vector.
Note that the implementation of the inequality
constraints usually results in distribution of solutions
with two modes at 0% and 100%. The increment is
calculated as a fraction of the distance on the concen-
tration scale between the median and the 100%
concentration. The incremental fraction is simply the
difference between the percentage frequencies at the
100% concentration of each surface.
After the ice concentration vector is determined from
the distribution of all possible solutions as being either
the median or incremented median, the concentrations
for the ice types and OW are normalized such that the
total concentration is 100%. The estimated concentra-
tion
c
k
is normalized by dividing by the summation of
the estimated concentrations of all surfaces
0
and
0.1
hh
vv
0
80-90
90-100
100-110 110-120
Radiometric values
120-130 130-140 140-150
1
Random number
generated
0.8
0.6
Selected
CRV
0.4
0.2
0
80
90
100
110
120
130
140
150
Radiometric values
Figure 10.13
Graphical illustration of the process to generate
a set of CRVs to mimic the probability distribution of a given
observation for a given ice type. An example of a coarse prob-
ability distribution is shown at the top panel and the corre-
sponding cumulative probability is shown by the broken line
at the bottom.
An example of the probability distributions of four
parameters
T
b,
36
h
and PR
36
from AMSR‐E 36.5 GHz
channel as well as the backscatter coefficients
and
0
from QuikSCAT Ku‐band are shown in Figure 10.14
for three ice types: YI, FY, and MY ice in addition to
open water. These parameters were used to generate
daily ice concentrations of the Arctic basin from 2002
to 2009 (when QuikSCAT was decommissioned). The
probability distributions were input to ECICE to gener-
ate the CRVs and eventually the concentration of each
ice type. The distributions should be generated using
samples from areas with homogeneous surface that can
be identified in the imagery data or using operational ice
charts. Anomalies caused by ice surface melt or wind‐
generated ocean surface can be incorporated with
appropriate weights to make the distribution as repre-
sentative of the ice types and the observed parameters
as possible. The overlap of the distributions in the figure
shows clearly the difficulty of using a single tie point to
represent an ice type uniquely; especially YI. It under-
scores the need for an optimization approach to deter-
mine the concentration of ice types, which is implemented
in ECICE. The higher backscatter of YI compared to
FYI probably results from the potentially rough surface
vv
n
c
j
.
j
1
The availability of the number of possible solutions allows
for the determination of a confidence level (CL) associated
with the final concentration answer. The CL is a measure
based on the mean absolute deviation (MAD). The latter is
defined as the average of the absolute values of the differ-
ence between the final solution concentration after normali-
zation and the concentration produced by each trial with a
different set of CRV (in percentage). The CL value (between
0 and 1) is derived from MAD according to the equation
