We use U-Net to detect fracture locations from satellite imagery across Antarctica (125m-resolution MOA imagery (2009); https://doi.org/10.7265/N5KP8037). Ice fractures result in the collapse of Antartctica ice-shelves, which can accelerate glacier flows into the ocean. We trained the U-Net with the labeled imagery, and applied the trained model to detect fracture across the Antarctic ice shelvse. In Fig. 1 the fracture/non-fracture locations are marked in white/black. The neural-network predicted continent-wide fracture locations (Fig. 1) is can be downloaded at https://doi.org/10.15784/601395, and viewed on the Google Earth Engine via https://charlottelai007.users.earthengine.app/view/moa2009fractureloc125m.
The Tensorflow U-Net implementation was developed by Akeret et al. (2017) and available at https://github.com/jakeret/tf_unet.
Input images and the corresponding labeled images are in the format of .tif. The filenames of the labeled images ends with "_mask.tif". The imput image is a grey-valued image with one channel. labeled images contains only two grey values corresponding to two two classes (fracture: 255 and non fractutre: 0).
-
data_trainset: training data (26 1000x1000 pixel tiles, shown in blue in Fig. 1)
-
data_validset: validation data (6 1000x1000 pixel tiles, shown in red in Fig. 1)
-
data_testset: testing data (6 1000x1000 pixel tiles, shown in green in Fig. 1)
-
data_ross: two extra images on the Ross ice shelf for visulization. One tile with fracture one without, the prediction of the fracture image is shown in Fig. 2.
-
Fracture_prediction_demo.ipynb: example code for training and testing UNet
C. Y. Lai, J. Kingslake, M. Wearing, P.-H. Cameron Chen, P. Gentine, H. Li, J. Spergel, J. M. van Wessem, “Vulnerability of Antarctica’s ice shelves to meltwater-driven fracture," Nature, 584, 574–578 (2020). doi: 10.1038/s41586-020-2627-8