12 Towards Operational Monitoring of Arctic Sea Ice by SAR
cient to be representative of the population of regions from each particular class. The
classification is carried out by displaying the underlying image together with an overlay of region boundaries (from the segmentation). The operator then selects regions
and assigns a class (effectively as an additional region attribute). This information is
added to the vector database. Once sufficient regions have been selected, the updated
vector database is written to disk as a vector training database.
In the second stage of the classification process, the training database is accessed in
a classification of the entire image, using the MLP or K-NN algorithm. This is a partially automated stage of the classification process and may involve more than one iteration as the resulting image classification may need to be improved by trying a new
classifier or, more likely, by refining the training database.
Typical results of the classification are shown in Fig. 3. From the original ERS-l SAR
image (Fig. 3a) acquired in the Kara Sea in March 1994, a classification product (Fig.
3b) and a total ice concentration product (Fig. 3C) are derived. Other products (not
shown) which can be generated by the system include an" old ice" concentration map
and a classification confidence map (based on the relative distance of attributes in each
region from the cluster center of its assigned class).
The region attributes used in the classification include both texture and mean
backscatter attributes, calculated using pixels located away from the region boundaries.
The selection is not based on an understanding of the relative value of each attribute,
although there is some evidence for benefits in the use of some texture measures
including the normalized standard deviation crill and inertia and entropy measures
from the spatial gray level dependency matrix (Haralick et al. 1973; Sun et al. 1992).
Instead, a range of attributes are included in a black box approach to maximize the
probability of including relevant attributes in each particular circumstance. By including the slant range coordinate of a region as an attribute, the classification can also take
into account incidence angle variations in the attributes (particularly in mean
backscatter). By providing scatter plots for the various attributes, and estimating their
separability from manually classified regions, it was found that crill and entropy were
useful discrimination measures for winter ice, whilst inertia, uniformity, and dissimilarity also added discrimination capability, but of more marginal value.
The hybrid classification approach advocated here is considered to be a reasonable compromise between an automated scheme which is unlikely to be valid outside very restricted circumstances and a fully manual scheme which fails to make use of any computer intelligence in assisting the user. Table 2 illustrates the division of classification tasks into manual and automated. Note in Table 2 that for the tasks specified as manual, it is not envisaged that the selection of the attributes would need to be performed on every occasion.
Figure 4 shows a comparison of AVHRR-derived ice concentrations and ERS-l SARderived ice concentrations for the Greenland sea, in a campaign carried out in August
1994. The AVHRR-derived concentrations are estimated from Danish Meteorological
Institute (DMI) ice charts recorded on the same day as each of eight ERS-l SAR images.
The errors on the AVHRR-derived ice concentrations result from the binning used by DMI
in producing the ice charts, and in the SAR-derived ice concentrations represent ±1 standard deviation. It can be seen that open water and pack ice are recognized consistently
using the two sensors. At moderate ice concentrations, there is a discrepancy of up to
2hoths, which to date has not been explained. In the long term, the availability of spaceborne SAR data with dual polarization will reduce the reliance on manual intervention.
cient to be representative of the population of regions from each particular class. The
classification is carried out by displaying the underlying image together with an overlay of region boundaries (from the segmentation). The operator then selects regions
and assigns a class (effectively as an additional region attribute). This information is
added to the vector database. Once sufficient regions have been selected, the updated
vector database is written to disk as a vector training database.
In the second stage of the classification process, the training database is accessed in
a classification of the entire image, using the MLP or K-NN algorithm. This is a partially automated stage of the classification process and may involve more than one iteration as the resulting image classification may need to be improved by trying a new
classifier or, more likely, by refining the training database.
Typical results of the classification are shown in Fig. 3. From the original ERS-l SAR
image (Fig. 3a) acquired in the Kara Sea in March 1994, a classification product (Fig.
3b) and a total ice concentration product (Fig. 3C) are derived. Other products (not
shown) which can be generated by the system include an" old ice" concentration map
and a classification confidence map (based on the relative distance of attributes in each
region from the cluster center of its assigned class).
The region attributes used in the classification include both texture and mean
backscatter attributes, calculated using pixels located away from the region boundaries.
The selection is not based on an understanding of the relative value of each attribute,
although there is some evidence for benefits in the use of some texture measures
including the normalized standard deviation crill and inertia and entropy measures
from the spatial gray level dependency matrix (Haralick et al. 1973; Sun et al. 1992).
Instead, a range of attributes are included in a black box approach to maximize the
probability of including relevant attributes in each particular circumstance. By including the slant range coordinate of a region as an attribute, the classification can also take
into account incidence angle variations in the attributes (particularly in mean
backscatter). By providing scatter plots for the various attributes, and estimating their
separability from manually classified regions, it was found that crill and entropy were
useful discrimination measures for winter ice, whilst inertia, uniformity, and dissimilarity also added discrimination capability, but of more marginal value.
The hybrid classification approach advocated here is considered to be a reasonable compromise between an automated scheme which is unlikely to be valid outside very restricted circumstances and a fully manual scheme which fails to make use of any computer intelligence in assisting the user. Table 2 illustrates the division of classification tasks into manual and automated. Note in Table 2 that for the tasks specified as manual, it is not envisaged that the selection of the attributes would need to be performed on every occasion.
Figure 4 shows a comparison of AVHRR-derived ice concentrations and ERS-l SARderived ice concentrations for the Greenland sea, in a campaign carried out in August
1994. The AVHRR-derived concentrations are estimated from Danish Meteorological
Institute (DMI) ice charts recorded on the same day as each of eight ERS-l SAR images.
The errors on the AVHRR-derived ice concentrations result from the binning used by DMI
in producing the ice charts, and in the SAR-derived ice concentrations represent ±1 standard deviation. It can be seen that open water and pack ice are recognized consistently
using the two sensors. At moderate ice concentrations, there is a discrepancy of up to
2hoths, which to date has not been explained. In the long term, the availability of spaceborne SAR data with dual polarization will reduce the reliance on manual intervention.