Exploring Frigid Frontiers: Strategies for Uncovering Earth's Frozen Regions With AI

1.1 U-Net: A CNN Based Architecture

In the field of computer vision a convolutional network design known as U-Net architecture is employed for segmentation of image task. U-Net initially was introduced by Olaf Ronneberger (Ronneberger et al., 2015) and has been applied in various application of image segmentations like medical image analysis. This research employs U-Net architecture for remote sensing data, which perform interpretation of satellite or aerial images.

Figure 1.

Area of study around continent Antarctica

(Baumhoer et al., 2019b)

The data set was taken from two satellite Sentinel1-1A (2014) and sentinel1-1B (2016) that have radar imaging sensors. Moreover 3-D images also taken from TanDEM-X (2013-2014) which is a German satellite for earth observation.

U-net architecture contains both an encoder and decoder and resembles to U -shaped structure. Its design mainly composed of contracting path(encoder) and expansive part (decoder) to provide precise object segmentation within images. This study uses a classifier that takes both pixel values and spatial context to make a distinction between land ice and ocean. It also makes few modifications to U-net architecture to meet its requirement. (Baumhoer et al., 2019b) The main modifications are: employing larger input tiles; initiating with 32 feature channel unlike 64 and scaling is done with 512 unlike 1024; incorporating dropout mechanisms; utilizing four input channels, as opposed to a single channel as described in Figure 2.

Figure 2.

It is a representation of U-Net Architecture where red arrows show down-sampling block, green arrows show up-sampling, black arrows show skip connections and yellow arrows indicate dropout

(Baumhoer et al., 2019b)

Figure 3.

Sequence of timeline around Getz Ice Shelf from Sentinel-1. (Baumhoer et al., 2019b) (A) Forward movement of the DeVicq Glacier. (B)Settled coastline part. (C) Breaking of ice chunks from front of Beakley Glacier (D)Wrong boundary prediction of Getz 1 Glacier front (E) Breaking of ice from Glacier event

Figure 3 captures the movement of DeVicq Glacier accurately as shown by time series. The parameters used for classification accuracy are given in following table. Accuracy varied across areas which are summarised in table 1.

Calculation of breaking of ice front of glacier around Antarctica will be allowed by the spatial and temporal transferability of this approach in future. This approach will provide new ways to monitor the changes in the Antarctica coastal region and can be used for measuring glacier and ice front extent fluctuations.