Geology Reference
In-Depth Information
The simple approach of using a set of thresholds to
separate ice types in a given parameter space (single or
multidimensional) was used in many studies to classify
ice types in single‐ or multichannel remote sensing
imagery data. Nevertheless, the overlap between radio-
metric signatures from ice types may render the selection
of the thresholds difficult and the classification results
sensitive to the selection. Furthermore, the technique
may not be robust enough, i.e., its validity is limited to the
set from which the thresholds have been selected. An
alternative approach entails using statistics of radiomet-
ric parameters from ice types as “training” data sets to
established classes in a given radiometric parameter space
using a supervised classification approach. This approach
requires also the selection of an appropriate metric and
assigning a given observation to the class that has its
center satisfying the condition of the minimum metric
(measured from the given observation to the center of
the established class). In an unsupervised classification
approach, on the other hand, the distribution of the
observations are examined in a given parameter space to
find if it leads to separable clusters. The unknown pixel
(i.e., the observation) is then assigned to the appropriate
cluster, also based on the concept of the minimum of a
selected metric. No training data are needed for this
approach but the resulting clusters are not labeled.
A well‐known and widely used probabilistic approach
of image classification is the Bayesian parameter estima-
tion or maximum‐likelihood estimation (these are nearly
identical methods but the approaches are slightly differ-
ent). The approach is based on statistical properties of
the radiometric measurements from each ice class. For a
set of satellite observations
R
the conditional probabil-
ity of occurrence of ice type ice
j
(including seawater as a
possible type) given an observation
R
o
is obtained from
the equation
10.1.1. Ice Classification from Optical and TIR Systems
Although the ice classes required for operational ice
monitoring encompass all the types based on stage of
development (Table 2.4), the climate‐related applications
require classification based on the three major age‐based
ice types; namely young ice (YI) (including new ice
types), FY ice, and MY ice. The two parameters that can
be derived from optical and TIR sensors and used in ice
classification are albedo and surface temperature, respec-
tively. Ice types that are sensitive to either or both param-
eters can be identified in the imagery data. The types that
can be identified based on their albedo are shown in
Table 8.9. A difficulty in using albedo for this purpose
is the significant deviation of its value from the typical
value of the given ice type when snow cover exists. This
is particularly true if snow metamorphoses or acquires
wetness. On the other hand, the major three categories
of ice types: young, FY, and MY ice can be identified
using TIR observations based on their presumably dif-
ferent surface temperature. The temperature is different
because the amount of heat exchange between the ocean
and the atmosphere is influenced by the ice thickness,
which is remarkably different between these three types.
Subcategories within any major ice type (e.g., FY ice
types based on thickness or surface deformation) do not
render unique albedo or surface temperatures that can
assist in their classification from optical or TIR data.
However, unlike the suite of ice classes required for
operational ice monitoring, the classification into the
above‐mentioned three major ice categories is useful for
climate‐related applications. Additionally, optical and
TIR data can be used to discriminate between ice and
O W.
Yu and Rothrock
[1996] found that the freezing
temperature of Arctic seawater varied between 271.36
and 271.45 K. A threshold of 271.4 K is usually used to
separate water and sea ice.
An early study to classify ice types from AVHHR visi-
ble and TIR data was conducted by
Massom and Comiso
[1994]. They used the NIR channel 2 (0.912
μ
m) and the
TIR channel 4 (10.8
μ
m) to discriminate between four
surfaces that existed in the Bering and Greenland Seas in
April and May 1988: OW, new ice, young ice (including
light Nilas), and FY ice (intermingled with some MY ice
advected from the central Arctic). The physical basis for
the ice classification was the different albedos and surface
temperatures (combined) between ice types and their val-
ues that contrast the value from OW. The frequency dis-
tribution of the data from channels 2 and 4 for the four
categories are shown in Figure 10.1. The albedo values
of OW and new ice are not distinguishable in the channel
2 data, but their different surface temperatures trigger
distinguishable signatures of brightness temperature
from the channel 4 data (recall the equal and narrow
p R
p
ice
p R
|
ice
j
j
(10.1)
ice
|
j
0
N
p
ice
p R
|
ice
k
k
k
1
where
p
(
R
|ice
j
) is the probability of occurrence of
R
given the ice type ice
j
,
p
(ice
j
) is the a priori probability of
ice
j
, and
N
is the number of ice types. In the case where
no prior knowledge or assumptions are available,
p
(ice
j
)
can be set to the inverse of the number of ice types and
OW; i.e., if the purpose is to assign the observation to
one of three possible surface types, then all
p
(ice
j
) can be
assumed one‐third for each type. The above equation
can be applied with any number of observations from
multisensor sources. The left‐hand side is determined
and the pixel is assigned to the class ice
j
that has the
highest probability.
