12 Towards Operational Monitoring of Arctic Sea Ice by SAR
12.2
Ice Classification and Concentration
12.2.1
Overview
There are three main approaches to the classification of sea ice in SAR data. The first
involves a purely manual classification, perhaps using a customized PC tool for producing ice charts. This approach makes the reasonable assumption that the human
interpreter is himself the best image analysis "tool" available. The second approach
is to use an automated classification, with image analysis tools classifying images
based on a predetermined set of optimized algorithms and algorithm parameters.
This technique has the advantage of minimizing processing time and user intervention. An example of this approach is that provided by the Alaskan SAR Facility geophysical processor system, where ERS-derived ice products make use of look-up
tables to classify image data (Kwok et al. 1992). A third, hybrid classification involves
manual classification of selected regions in an image and automated (but supervised) classification of the remainder of the regions based on their similarity to the
manually classified regions. An example of this approach is provided by the IPAP
image analysis tools.
An ideal classification tool would be fully automated, to minimize processing time
and the requirement for skilled labor, but there are two main reasons why this is difficult to implement at the current time, namely the problems of speckle and inherent
variability in backscatter properties.
12.2.1.1
Speckle
SAR data are subject to "speckle;' which is noise inherent to the coherent imaging
process. For intensity data from a homogeneous region, multilook speckle is gamma-distributed with:
cr/ll=l/ffi
where 11 is the mean pixel amplitude, cr is the local standard deviation, and N is the
effective number of independent looks. The number of looks is determined by the processing bandwidth and any smoothing after image construction, and can be estimated
from the window size used to calculate cr and the spatial resolution of the data. Speckle will degrade any estimate of the backscatter coefficient and hence our ability to classify an image based on this parameter alone.
If the data are sufficiently smoothed, then speckle is reduced, but this is achieved at
the expense of poor spatial resolution (see Sect. 12.2.2). If higher spatial resolution is
required, then it is necessary to carry out segmentation of the image data into homogeneous regions of ice and open water prior to classification. By extracting attributes
of regions rather than individual pixel intensities, the effect of noise (particularly on
estimates of backscatter coefficient) is reduced. Importantly, segmentation also allows
attributes other than mean backscatter to be measured, and this can assist in region
12.2
Ice Classification and Concentration
12.2.1
Overview
There are three main approaches to the classification of sea ice in SAR data. The first
involves a purely manual classification, perhaps using a customized PC tool for producing ice charts. This approach makes the reasonable assumption that the human
interpreter is himself the best image analysis "tool" available. The second approach
is to use an automated classification, with image analysis tools classifying images
based on a predetermined set of optimized algorithms and algorithm parameters.
This technique has the advantage of minimizing processing time and user intervention. An example of this approach is that provided by the Alaskan SAR Facility geophysical processor system, where ERS-derived ice products make use of look-up
tables to classify image data (Kwok et al. 1992). A third, hybrid classification involves
manual classification of selected regions in an image and automated (but supervised) classification of the remainder of the regions based on their similarity to the
manually classified regions. An example of this approach is provided by the IPAP
image analysis tools.
An ideal classification tool would be fully automated, to minimize processing time
and the requirement for skilled labor, but there are two main reasons why this is difficult to implement at the current time, namely the problems of speckle and inherent
variability in backscatter properties.
12.2.1.1
Speckle
SAR data are subject to "speckle;' which is noise inherent to the coherent imaging
process. For intensity data from a homogeneous region, multilook speckle is gamma-distributed with:
cr/ll=l/ffi
where 11 is the mean pixel amplitude, cr is the local standard deviation, and N is the
effective number of independent looks. The number of looks is determined by the processing bandwidth and any smoothing after image construction, and can be estimated
from the window size used to calculate cr and the spatial resolution of the data. Speckle will degrade any estimate of the backscatter coefficient and hence our ability to classify an image based on this parameter alone.
If the data are sufficiently smoothed, then speckle is reduced, but this is achieved at
the expense of poor spatial resolution (see Sect. 12.2.2). If higher spatial resolution is
required, then it is necessary to carry out segmentation of the image data into homogeneous regions of ice and open water prior to classification. By extracting attributes
of regions rather than individual pixel intensities, the effect of noise (particularly on
estimates of backscatter coefficient) is reduced. Importantly, segmentation also allows
attributes other than mean backscatter to be measured, and this can assist in region