These observations indicate that higher level image features
(such as compositions of basic image structures such as edges
and corners) need to be used in order to capture ice appearances
in SAR imageries under variable and changing surface conditions.
The design of such features is difficult and nonintuitive.
Redesign of the algorithm or features may be necessary for
different regions or times of the year. Therefore, it is of interest
to investigate the use of feature learning instead of feature
design for ice concentration estimation from SAR images. If
features can be learned from training data, the most effective
features can be generated for specific data and conditions
without human intervention. This is highly desirable for largescale
automatic ice concentration estimation tasks.