Geology Reference
In-Depth Information
camera survey, and in situ measurements and observa-
tions. Results from ice motion algorithms that identify
divergence and convergence of the ice cover have also
been used as a source of validation of results from ice
concentration algorithms, yet with some limitations
[
Kwok
, 2002]. The operational ice centers have the most
stringent requirements to the accuracy of the charts, but
they are generated at a coarse resolution (a few kilometers
or tens of kilometers). Centers that produce these charts
are listed in Table 10.2. Visual observations by marine
operators from ships are not always available and can be
limited by the perimeter of validity of the information
around the ship due to the grazing angle of the view.
Airborne observations cover wide areas but they are
expensive. The most reliable source of validation data is
field measurements and observations. However, these
data are not always available especially in the polar
regions because of the long periods of darkness and
cloud cover. Field expeditions are also expensive.
Moreover, in situ point measurements may not be repre-
sentative of the large footprint of the satellite observa-
tion. The conclusion is that except for extensive field
measurements or just visual observations of the ice cover,
no data set should be qualified as being the “truth.”
Information in operational ice charts, which are com-
monly used as a source for validating the calculated ice
concentration and other parameters, are subjectively pro-
duced by ice analysts as they delineate areas of uniform
spatial distribution of ice types and concentrations (sec-
tion 11.2). If the footprint of the sensor from at which the
ice concentration is calculated is much smaller than the
size of the polygon, then a discrepancy between the algo-
rithms results and the ice concentration assigned to the
polygon should be expected. In other words, the compari-
son will reveal the heterogeneity of the ice concentration
within a polygon, which is supposed to have a uniform
distribution of ice types (hence concentration). This
point will be illustrated further with an example in this
subsection.
Aside from the difficulty of obtaining truth data,
remote sensing observations suffer from a few inherent
errors that must be taken into consideration during the
validation process. These include anomalies of radiomet-
ric measurements (as discussed in the previous section)
and inaccuracies of geolocations of pixels.
Anderson
[2000] identified three criteria for the evalu-
ation of ice concentration algorithms: the accuracy
of concentration estimate, the noise resistance, and the
resolution at which the maps are produced. Prior to using
these or any other criteria, the validation data should be
provided at a scale comparable to the spatial resolution
of the retrieved concentration. Given the ensuing errors
in any set of validation data, a best recommendation to
validate the retrieved ice parameters would be to use as
many sources of validation data as possible in order to
reach a reasonable conclusion about the domain of appli-
cability of the results. This approach has been used in a
few studies of intercomparison of ice concentration from
different algorithms and different sensors [e.g.,
Ivanova
et al
., 2014] as shown in the next section.
An approach to validate ice concentration results
against estimates from an operational image analysis of
Radarsat imagery data in the CIS has been developed by
Shokr and Moucha
[1998]. In addition to the validation
of the given ice concentration algorithm, it can be used
to evaluate the information from the operational image
analysis and ice charts. This approach can be categorized
as a data fusion (or data co‐location) technique. The rest
of this section addresses the description of the technique
with results from the evaluation of the ECICE
algorithm.
The method maps footprints of a coarse‐resolution
sensor (e.g., a passive microwave) onto an image from
a fine‐resolution sensor (e.g., Radarsat). An example is
shown in Figure 10.20 where footprints from SSM/I
37 GHz channel are overlaid onto a near‐coincident
Radarsat‐1 subscene. Each SSM/I footprint contour is
established as a set of latitude‐longitude coordinates of
12 vertices (dodecagon) that approximate its elliptical
shape. They are calculated using the latitude/longitude
coordinates of the SSM/I pixel (the center of the foot-
print), along with the dimensions of the major and minor
axes of the footprint. A simple spherical Earth surface
model is also used to determine the orientation of the
footprint [
Shokr
, 2004]. The spatial accuracy of the co‐
located footprints is determined by the geolocation accu-
racy of pixels from each imagery data set. Typically, the
accuracy is 2-5 pixel spacing for a fine‐resolution sensor
such as Radarsat and one pixel from a coarse‐resolution
sensor such as SSM/I.
Figure 10.21 shows how the data co‐location method
can be used to compare total ice concentration products
from ECICE and NT2 algorithms (using SSM/I) against
the corresponding Radarsat‐1 image analysis chart from
the CIS. The scene is of the southern part of the Gulf of
St. Lawrence and its exit to the Atlantic Ocean, acquired
on 18 March 2003. The Radarsat image is shown with the
delineated CIS Radarsat image analysis polygons (IAPs)
overlaid (see section 11.2). The polygons are colored in
panels (b), (c), and (d) to reflect the subjectively estimated
ice concentrations according to the attached color bar.
Most of the polygons were estimated to have 100% ice
concentration. A few polygons at the bottom right corner
of the scene had concentrations between 20% and 70%.
The elliptical footprints of the SSM/I 85 GHz channel are
displayed in panel (b) and of SSM/I 37 GHz channel in
panels (c) and (d). The footprints are colored to represent
the calculated ice concentrations from NT2 in panel (c)
