This study for automatically mapping glacier calving fronts was inspired by observing how a human would annotate land and glacier areas in a satellite image:

During the annotation process, thoughts like the following might run through the annotators head:

1. “Good thing I can zoom in and out as needed”
2. “I’ll start with the edge”
3. “Give me a bigger brush!”

The proposed HED-UNet model was designed by incorporating these ideas.

HED-UNet combines the commonly used semantic segmentation model UNet with the edge detection model HED to reflect the need to focus on the edges.

For allowing the model to use a larger spatial context and “use a bigger brush”, the number of up- and downsampling layers was increased from 4 to 6. Finally, to retain fine-grained details near the edges but use broader contextual information farther away from the edges, the predictions at different resolution levels are merged through an attention merging head.

As a multi-task model, HED-UNet predicts both segmentation maps and edge detection maps. Compared to state-of-the-art models for both these tasks, it shows superior performance when applied for calving front detection.

Inspecting the attention maps used for merging information from different resolutions shows that the model is indeed behaving as expected: High-resolution information is used in the boundary areas, while the more robust, low-resolution information is used in other regions.

The full paper is available on IEEE Explore

If you would like to have a closer look at the implementation details, work with our method, or reproduce our results, you can find all of our code on Github .