94
S.G. BEAVEN AND S.P. GOGINENI
In addition to the basic architectures outline above, it is possible to develop a technique that is actually a hybrid of two of these. Data fusion is thought to be most beneficial if it is performed "as close to the sensor as possible;' although the level at which
fusion is performed depends on the nature of the problem and the characteristics of
the sensors to be used or developed.
5.3
Satellite Passive Microwave Algorithms
Satellite microwave radiometers have been used for estimating the fractional coverage
of sea ice nearly continuously since the 1970S (Carsey 1982; Gloersen et al. 1992). The
NT algorithm and the bootstrap algorithm have been used to determine sea ice concentration and type in the Arctic (Cavalieri et al.1984; Comiso 1986; Gloersen and Cavalieri 1986; Gloersen et al. 1992). The NT algorithm was originally developed to determine sea ice concentration from daily averaged Scanning Multichannel Microwave
Radiometer (SMMR) microwave brightness temperatures over the Arctic and has been
adapted to DMSP SSM!I data. It is based on the differences in the emissivity characteristics of multiyear ice, first-year ice and open water at 37 GHz and 19 GHz and the
difference between vertically and horizontally polarized emissivity at 19 GHz. It was
developed primarily for use during winter months.
The mean difference between the estimates of total ice concentration during fall (late
September through November) obtained using the NT algorithm and from the Landsat imagery is less than 1%, with a standard deviation of ±7.4% (Steffen and Schweiger
1991). The results for fall are indicative of the performance of the algorithm in winter
because of similar ice conditions. Steffen and Schweiger (1991) have also shown that the
mean difference between the NT and Landsat-based estimates of total ice concentration is 2.4% during spring, with a standard deviation of ±5.9%. These results were
obtained using hemispheric tie points; they have shown that some improvements can
be made through the use of local tie points.
During the summer melt period, however, ice concentration estimates from the NT
algorithm differ from Landsat estimates by as much as 46%, with a mean difference of
11% (Steffen and Schweiger 1991). In addition, the NT algorithm underestimates multiyear ice concentration during fall and through the early winter. When air temperatures are around 0 0 C, it overestimates first-year ice concentration and underestimates
multiyear ice concentration (Carsey 1985; Cavalieri et al. 1990). This effect continues
into the onset of freeze-up.
Using SMMR data, Rothrock and Thomas (1992) have shown that multiyear ice concentration in the central Arctic just after freeze-up is inconsistent with the total ice concentration at the end of the summer melt. The NT algorithm yields a multiyear ice concentration of about 0.4 in the central Arctic, whereas the total ice concentration at the
end of summer was about 0.7. Thomas (1993) has shown that the multiyear concentration estimates from the NT algorithm reach a minimum during September, and this
minimum is not believed to be caused by a real decrease in multiyear ice concentration, but by changes in the signature of multiyear ice. The underestimation of multiyear ice is attributed to the use of constant pure ice type signatures in the NT algorithm.
He has shown that by accounting for this variability through the use of daily mean signatures, estimates of multiyear concentration may be improved. This appears to be asso-
S.G. BEAVEN AND S.P. GOGINENI
In addition to the basic architectures outline above, it is possible to develop a technique that is actually a hybrid of two of these. Data fusion is thought to be most beneficial if it is performed "as close to the sensor as possible;' although the level at which
fusion is performed depends on the nature of the problem and the characteristics of
the sensors to be used or developed.
5.3
Satellite Passive Microwave Algorithms
Satellite microwave radiometers have been used for estimating the fractional coverage
of sea ice nearly continuously since the 1970S (Carsey 1982; Gloersen et al. 1992). The
NT algorithm and the bootstrap algorithm have been used to determine sea ice concentration and type in the Arctic (Cavalieri et al.1984; Comiso 1986; Gloersen and Cavalieri 1986; Gloersen et al. 1992). The NT algorithm was originally developed to determine sea ice concentration from daily averaged Scanning Multichannel Microwave
Radiometer (SMMR) microwave brightness temperatures over the Arctic and has been
adapted to DMSP SSM!I data. It is based on the differences in the emissivity characteristics of multiyear ice, first-year ice and open water at 37 GHz and 19 GHz and the
difference between vertically and horizontally polarized emissivity at 19 GHz. It was
developed primarily for use during winter months.
The mean difference between the estimates of total ice concentration during fall (late
September through November) obtained using the NT algorithm and from the Landsat imagery is less than 1%, with a standard deviation of ±7.4% (Steffen and Schweiger
1991). The results for fall are indicative of the performance of the algorithm in winter
because of similar ice conditions. Steffen and Schweiger (1991) have also shown that the
mean difference between the NT and Landsat-based estimates of total ice concentration is 2.4% during spring, with a standard deviation of ±5.9%. These results were
obtained using hemispheric tie points; they have shown that some improvements can
be made through the use of local tie points.
During the summer melt period, however, ice concentration estimates from the NT
algorithm differ from Landsat estimates by as much as 46%, with a mean difference of
11% (Steffen and Schweiger 1991). In addition, the NT algorithm underestimates multiyear ice concentration during fall and through the early winter. When air temperatures are around 0 0 C, it overestimates first-year ice concentration and underestimates
multiyear ice concentration (Carsey 1985; Cavalieri et al. 1990). This effect continues
into the onset of freeze-up.
Using SMMR data, Rothrock and Thomas (1992) have shown that multiyear ice concentration in the central Arctic just after freeze-up is inconsistent with the total ice concentration at the end of the summer melt. The NT algorithm yields a multiyear ice concentration of about 0.4 in the central Arctic, whereas the total ice concentration at the
end of summer was about 0.7. Thomas (1993) has shown that the multiyear concentration estimates from the NT algorithm reach a minimum during September, and this
minimum is not believed to be caused by a real decrease in multiyear ice concentration, but by changes in the signature of multiyear ice. The underestimation of multiyear ice is attributed to the use of constant pure ice type signatures in the NT algorithm.
He has shown that by accounting for this variability through the use of daily mean signatures, estimates of multiyear concentration may be improved. This appears to be asso-