S Fusion of Satellite SAR with Passive Microwave Data for Sea Ice Remote Sensing
97
al effort to assign the result oflocal thresholding to each particular ice type. The technique, as used in our multisensor fusion process, is described below.
Local area thresholding is based on computation of histograms over small, overlapping regions of an image that are assumed to be small enough to contain only two distinct classes based on image intensity. The threshold is established on the basis of estimates of the distributions of the two classes in each subarea. Thresholds based on local
area statistics are used since local changes in mean backscatter often occur within SAR
image frames. These local changes in backscatter cause fluctuations in image intensity, which may cause problems in global thresholding techniques. We used subregions
that were 32 x 32 pixels, with 50% overlap, from low-resolution ERS-l SAR imagery. Histograms computed from entire ERS-l SAR image frames do not appear to be bimodal,
as shown in Fig. 3a, for example. However, subregions of this same image yield distributions that appear to be bimodal, as shown in Fig. 3b. The local area histogram shows
a bimodal distribution over the subregion, which is not apparent from the global histogram. This demonstrates that although no clear threshold may be obtained from a
global distribution of SAR data, local distributions may show distinct classes. Dynamic thresholding is based on this premise and is described in detail in Haverkamp et al.
(1995).
We used the DLT algorithm to separate each SAR image frame into three classes. These
classes are not defined by the method, and must be assigned to ice type through further processing or through inspection by an expert. To simplify the study of multisensor fusion for ice classification we identified which category corresponds to multiyear
ice through visual identification of multiyear floes within the class. An automated
method could require use of shape or other information as described in Haverkamp et
al. (1995). The multiyear ice concentration is then estimated over areas defined by SSM/I
grid boundaries. These boundaries were determined by the colo cation methods discussed below.
5.4.2
Active/Passive Image Registration
The use of multisensor data first requires registration or mapping of the data to a common coordinate system. When the two sensors have drastically different resolution it
is best to map the low-resolution data to the same coordinates as the high-resolution
data. We remapped the SSM/I brightness temperature grids to the ERS-l SAR image
frame coordinates. The process consists of computing the latitude/longitude pairs for
each pixel in the SAR image, then determining the corresponding pixel location within the SSM/I grids for the Arctic. These are stored in look-up tables for mapping the
SSM/I pixel locations to the SAR image frame. A summary of both steps is given below.
The geographic locations of the corner points of each SAR image frame are supplied
by the ERS-l Processing and Archive Facilities (PAFs) in the header file for each ERS-l
SAR image. We use these corner points to derive the latitude and longitude coordinates
for each pixel within the image using the algorithm given by Li (1994). This algorithm
uses the fact that ERS-l SAR imagery covers a small spatial area, eliminating the need
for using the more general formulation given by Curlander and MCDonough (1992).
The transformation from ERS-l SAR image coordinates to geographical coordinates
assumes that the boundaries of the image follow the great circle on the Earth. This
97
al effort to assign the result oflocal thresholding to each particular ice type. The technique, as used in our multisensor fusion process, is described below.
Local area thresholding is based on computation of histograms over small, overlapping regions of an image that are assumed to be small enough to contain only two distinct classes based on image intensity. The threshold is established on the basis of estimates of the distributions of the two classes in each subarea. Thresholds based on local
area statistics are used since local changes in mean backscatter often occur within SAR
image frames. These local changes in backscatter cause fluctuations in image intensity, which may cause problems in global thresholding techniques. We used subregions
that were 32 x 32 pixels, with 50% overlap, from low-resolution ERS-l SAR imagery. Histograms computed from entire ERS-l SAR image frames do not appear to be bimodal,
as shown in Fig. 3a, for example. However, subregions of this same image yield distributions that appear to be bimodal, as shown in Fig. 3b. The local area histogram shows
a bimodal distribution over the subregion, which is not apparent from the global histogram. This demonstrates that although no clear threshold may be obtained from a
global distribution of SAR data, local distributions may show distinct classes. Dynamic thresholding is based on this premise and is described in detail in Haverkamp et al.
(1995).
We used the DLT algorithm to separate each SAR image frame into three classes. These
classes are not defined by the method, and must be assigned to ice type through further processing or through inspection by an expert. To simplify the study of multisensor fusion for ice classification we identified which category corresponds to multiyear
ice through visual identification of multiyear floes within the class. An automated
method could require use of shape or other information as described in Haverkamp et
al. (1995). The multiyear ice concentration is then estimated over areas defined by SSM/I
grid boundaries. These boundaries were determined by the colo cation methods discussed below.
5.4.2
Active/Passive Image Registration
The use of multisensor data first requires registration or mapping of the data to a common coordinate system. When the two sensors have drastically different resolution it
is best to map the low-resolution data to the same coordinates as the high-resolution
data. We remapped the SSM/I brightness temperature grids to the ERS-l SAR image
frame coordinates. The process consists of computing the latitude/longitude pairs for
each pixel in the SAR image, then determining the corresponding pixel location within the SSM/I grids for the Arctic. These are stored in look-up tables for mapping the
SSM/I pixel locations to the SAR image frame. A summary of both steps is given below.
The geographic locations of the corner points of each SAR image frame are supplied
by the ERS-l Processing and Archive Facilities (PAFs) in the header file for each ERS-l
SAR image. We use these corner points to derive the latitude and longitude coordinates
for each pixel within the image using the algorithm given by Li (1994). This algorithm
uses the fact that ERS-l SAR imagery covers a small spatial area, eliminating the need
for using the more general formulation given by Curlander and MCDonough (1992).
The transformation from ERS-l SAR image coordinates to geographical coordinates
assumes that the boundaries of the image follow the great circle on the Earth. This