CALFIN: A Calving Front Mask Dataset for Greenland, 1972-2020

Abstract:

We present an update on Calving Front Machine (CALFIN), a dataset of calving front positions for glaciers in Greenland. The dataset provides automatically generated positions from Landsat imagery, from 1972 to June 2020. This dataset provides sub-seasonal constraints on glacial evolution for 80+ basins. Additions include coverage along North, Northeast, and Southern coasts, as well as data for 2019-2020. Thus this dataset offers the modeling community improved opportunity to explore previous trends and validate existing models.

CALFIN uses a deep neural network to automatically generate the calving fronts from Landsat imagery. This neural network uses a UNet-based DeeplabV3+ Xception architecture. This design builds on existing work from Mohajerani, Zhang, and Baumhoer. Additional post-processing techniques allow our method to achieve accurate, useful segmentation of raw images into masks and Shapefile outputs. This methodology is uniquely robust to clouds, illumination differences, sea ice, and Landsat 7 scan-line errors.

Lastly, we provide an error analysis of our results, and compare CALFIN’s performance against existing methodologies. We achieve a mean error of 2.25 pixels (86.76 meters) from the true front on a diverse set of 162 testing images, and anticipate the release of 20,000+ processed calving fronts with ~100m mean error.

At this stage, we continue to seek feedback from the community. We welcome any critiques or questions regarding the dataset and/or our methods. This work was conducted as a collaboration between NASA’s Jet Propulsion Laboratory and the University of California, Irvine.