In this paper, a CNN has been applied to dual-polarized SAR
(HH and HV) images to generate ice concentration estimates.
The CNN used takes the image patches of the intensity-scaled
dual-pol SAR images as input and outputs ice concentration
directly. Image analysis charts are used for training. State-ofthe-
art pixel-level result is acquired during the melt season in
the Beaufort Sea. The ice concentration from the CNN contains
abundant details of the ice compared with the image analyses.
The results suggest that CNN is a robust method that can model
the effect of incidence angle, SAR image noise, and the effect
of wind on water and melt. Low-ice-concentration regions are
also captured by the CNN model used.
The training on 11 images takes about 9 h, the prediction
of ice concentration for one image takes around 10 min using
an Nvidia GTX 780 graphic card with around 2000 processing
cores. With more powerful graphic cards and appropriate optimization,
the processing time can be largely reduced. Once
the model is trained, the ice concentration estimation using
the model can run in parallel on multiple graphics processing
units easily. This is very promising for operational applications
which require timely and robust ice concentration estimates
over large regions.
This is a case study due to the limited amount of data
available. For operational use, more comprehensive study and
evaluation of this method in different regions and times of the
year are required, which also requires the understanding of
the connection between the CNN features and the sea surface
conditions, which can be considered in future studies.