Automating iceberg detection in Greenland using deep learning on high to moderate-resolution optical imagery

Iceberg calving constitutes nearly half of the mass loss from the Greenland Ice Sheet, yet iceberg distribution around Greenland is poorly known. Icebergs act as a distributed source of fresh water to the ocean, reduce salinity, and affect fjord circulation. Therefore, quantifying the distribution of icebergs near the termini of outlet glaciers is important for understanding the solid flux of mass loss from ice sheets, as well as serving as a crucial input to global ocean models. The advent of remote sensing has greatly facilitated studying the distribution of icebergs. Since icebergs appear in a variety of shapes and sizes in polar regions, investigating their distribution requires robust automated tools. Several studies have attempted to automate iceberg detection; however, they often rely on threshold-based algorithms that introduce large segmentation errors. Therefore, there is a need for a reliable segmentation tool for iceberg detection.

Here we present spatio-temporal distribution of icebergs in Sermilik fjord west Greenland using high resolution optical Planet Imagery for summers of 2018 and 2019. We use a deep convolutional neural network, U-Net, to automatically identify the icebergs. The U-Net architecture consists of a downsampling and an upsampling branch for object identification and localization within the image. We train the network with 400 images and optimize the algorithm to learn the edges of neighboring icebergs properly. Comparing our results with two commonly-used segmentation algorithms confirms the superiority of deep learning methods to generic vision algorithms. Using transfer learning, we apply our trained model to Landsat and Sentinel-2 optical imagery and evaluate whether high (~3 m) resolution imagery is necessary to estimate iceberg distribution. Our approach can be used to provide iceberg distribution as a second-level product of the currently available optical imagery.