Classification of sea ice types on Sentinel-1 SAR data using convolutional neural networks

Hugo Bolze
ENSG, France
Seminar Date: 
23. August 2019 - 13:00 - 13:45
Lecture room, Ground Floor, NERSC

In order to assure the security of navigation and offshore activities scientists have to predict with precision the type and the location of sea ices. Getting information about cold intensity and ocean-atmosphere dynamics of these polar regions are also motivations. The ways to describe sea ices are numerous but this article focuses only on type.
The process developed in this article is based on Sentinel-1 A/B SAR images. Polar regions are often under bad weather conditions or without enough sunlight like during polar nights. It is impossible to use classic optic sensor to study sea ice this is why SAR technology is used.
Type of sea ices have already been classified through different approaches as developed in [Park et al., 2019] where sea ices are classified with Random Forests algorithm or manually with satellite data treated by an ice expert. However, scientists want to try another way to classify them with convolutional neural network (CNN). The main goals of this approach are to increase the current accuracy and to decrease the process time.