Geology Reference
In-Depth Information
algorithms return near 100% concentration in an area
delimited by daily ensemble mean concentration > 90%.
Usually all algorithms succeed in reproducing the 100%
ice concentration under cold polar winter conditions.
Shokr and Markus
[2006] compared results from AES‐
York and NT2 algorithms against results from the opera-
tional analysis of Radarsat‐1 images produced at the CIS.
Among four scientific questions addressed in the study
there was a question about the effect of the actual ice
concentration on the performance of each algorithm.
They found that the total ice concentration from the two
algorithms generally agree with CIS estimates at all con-
centrations. Another question was about the relative
concentration of thin and thick ice within the sensor's
footprint: would that ratio affect the estimate of the par-
tial concentration of each ice type? And if so, then how is
it affecting the results?
The authors found that the estimation of thin and
thick ice concentrations depends on the actual (true) ice
composition of ice in the footprint (thin ice here is
defined as ice < 15 cm thick). Figure 10.23 explains this
point. It shows the deviation of the retrieved ice concen-
tration from the corresponding CIS estimate versus the
difference between the thick and thin ice concentration
as determined by CIS. If thin ice dominates the footprint
area (negative values on the horizontal axis), then both
algorithms underestimate thin ice concentration with
respect to the CIS analysis (accordingly overestimate
thick ice types). On the other hand, if thick ice domi-
nates the footprint, then both algorithms underestimate
thick ice and accordingly overestimate thin ice concen-
tration (accordingly underestimate thick ice). The most
accurate retrieval of thin and thick ice concentration was
achieved when the footprint is composed of a balanced
mixture of thin and thick ice, regardless of the total ice
concentration.
Shokr and Kaleschke
[2012] compared results from
five algorithms NT, NT2, BS, ASI, and ECICE using
data from 100% ice concentration developed in an out-
door tank. The emitted radiation was measured using a
ground‐based radiometer so that the atmospheric influ-
ences were not present except for the reflected down-
welling radiation, which was corrected. The ice was thin
as it grew from 0 to 24 cm during the duration of the
experiment that lasted 48 days starting from the onset
of freezing on 23 November 2005. Detailed snow/ice
measurements were obtained to support interpretation
of the measured radiation and consequently the output
concentrations. Figure 10.24 shows the data obtained
from the first 6 days after the onset of freezing; days
328-333 (24 November to 29 November 2005).
Ice thickness reached 4 cm during this period with a
slushy surface most of the time. Within this thickness
range, most ice concentration retrieval algorithms fail to
estimate the correct value. The top panel in Figure 10.24
shows the evolution of the measured brightness tempera-
ture from the six radiometric channels, annotated with
atmospheric temperature, surface salinity, and composi-
tion as well as selected weather events. The middle panel
includes results from NT2, ASI, BSA, and NT algorithms
using appropriate tie points averaged from the measured
data. The bottom panel includes results from ECICE
using the three sets of input data from 19, 37, and 85 GHz
separately. Each set included observations from the hori-
zontal and vertical channels plus the polarization ratio.
The true concentration is 100%. Any deviation from this
value represents an error in the algorithm's estimate. The
NT, NT2, ASI, and BS underestimate the concentration
most of the time. ECICE, on the other hand, produces
the expected 100% concentration especially from using
the 37 or 85 GHz data. This is probably the advantage of
using a distribution of the radiometric parameters that
encompass all possible conditions of the given ice surface
instead of using a single tie point. For the algorithms that
use a single tie point, the data in Figure 10.24 were gener-
ated with sets of tie points that represent average of radi-
ometric parameters from experimental measurements.
It is interesting to note the cyclic pattern of the sharp
drop in
T
b
(e.g., observed at 329:07.01) followed by a
gradual increase. This is mimicked in the ice concentra-
tion estimates from all algorithms. This pattern is proba-
bly a manifestation of sudden increase of surface wetness
followed by gradual freezing. This phenomenon was
observed for ice with thickness <6 cm [
Shokr et al
., 2009].
A thin ice surface is usually covered with a layer of highly
saline brine formed due to vertical brine expulsion from
the subsurface layer as the ice temperature drops [
Weeks
&
Ackley
, 1982]. When snow falls on this surface, it melts
immediately, causing a significant increase in wetness and
decrease in the surface salinity. That explains the sudden
drop in brightness temperature, which is always observed
during the snow fall as labeled in Figure 10.24.
Recently,
Ivanova et al
. [2014] compared 11 algorithms
of sea ice concentrations. The algorithms are NORSEX,
NORSEX‐85, NT, UMass‐AES, BS, N90GHz, Cal/Val,
BRI, TUD, NT2, and ASI. The authors recognized the
difficulty of finding the “true” value of ice concentration,
so they used the average of all algorithms as the reference.
This is the only feasible approach to achieve at least rela-
tive evaluation of each algorithm. They found that the
algorithms tend to agree with each other more in the cen-
tral Arctic in winter when ice is mostly fully consolidated
and the atmosphere is less variable (differences are lim-
ited to around 2%). The largest deviation of results from
the different algorithms with respect to the reference
value in winter was between 5% and 12%. That was
mainly found in seas that feature low ice concentration
(e.g., the Sea of Okhotsk as well as the Bering, Labrador,
