In this paper, the estimation of ice concentration under the
framework of deep learning is investigated. Deep learning is
a feature learning method that uses multiple layers of neural
networks [17], which has recently demonstrated great potential
in many different recognition tasks such as speech recognition
and image object classification [18]–[21]. A deep convolutional
neural network (CNN) has been used to estimate ice concentration
directly from dual-band SAR images (HH and HV) in
the melt season. CNN is a neural network model that enforces
weight sharing and local connections between adjacent layers
of neurons. This method has a demonstrated ability to achieve
high performance for image-related recognition tasks [20],
[22]–[24]. In this paper, its capability to generate high-quality
ice concentration estimation in the melt season is demonstrated by a case study in the Beaufort Sea, under a variety of surface
conditions, without incidence angle correction or postprocessing.
A state-of-the-art ice concentration estimate from SAR
imageries is achieved. To the best of our knowledge, this is the
first study in which a CNN is used to extract ice concentration
from SAR images.