How to Get the Most Out of U-Net for Glacier Calving Front Segmentation

Abstract

The melting of ice sheets and glaciers is one of the main contributors to global sea-level rise. Hence, continuous monitoring of glacier changes and in particular the mapping of positional changes of their calving front is of significant importance. This delineation process, in general, has been carried out manually, which is time-consuming and not feasible for the abundance of available data within the past decade. Automatic delineation of the glacier fronts in synthetic aperture radar (SAR) images can be performed using deep learningbased U-Net models. This article aims to study and survey the components of a U-Net model and optimize the model to get the most out of U-Net for glacier (calving front) segmentation. We trained the U-Net to segment the SAR images of Sjogren-Inlet and Dinsmoore-Bombardier-Edgworth glacier systems on the Antarctica Peninsula region taken by ERS-1/2, Envisat, RadarSAT-1, ALOS, TerraSAR-X, and TanDEM-X missions. The U-Net model was optimized in six aspects. The first two aspects, namely data pre-processing and data augmentation, enhanced the representation of information in the image. The remaining four aspects optimized the feature extraction of U-Net by finding the best-suited loss function, bottleneck, normalization technique, and dropouts for the glacier segmentation task. The optimized U-Net model achieves a dice coefficient score of 0.9378 with a 20% improvement over the baseline U-Net model, which achieved a score of 0.7377. This segmentation result is further post-processed to delineate the calving front. The optimized U-Net model shows 23% improvement in the glacier front delineation compared to the baseline model.

Calving fronts and where to find them: a benchmark dataset and methodology for automatic glacier calving front extraction from synthetic aperture radar imagery

“…As our test set is out of sample (i.e., the study sites in the test set were not covered in the training set), the test set is entirely independent of the training set and, hence, qualifies for estimating the generalizability of our trained models to new unseen glaciers. Some previous calving front delineation studies (Cheng et al, 2021;Davari et al, 2022Davari et al, , 2021Hartmann et al, 2021;Holzmann et al, 2021;Marochov et al, 2021;Periyasamy et al, 2022;Zhang et al, 2019) used in-sample test sets (i.e., the test set includes images from the same glaciers as covered in the training set but from different time points). Deep learning models will produce more precise front delineations on in-sample test sets compared with out-of-sample test sets (Marochov et al, 2021), as the generalization gap between the training and test set is smaller for in-sample test sets (Quinonero-Candela et al, 2008).…”

“…Davari et al (2021) use a distance map-based loss to train their U-Net. Periyasamy et al (2022) further optimize the feature extraction of the U-Net. Lastly, Davari et al (2022) formulate the front segmentation as a regression task, letting a U-Net predict the distance of each pixel to the front.…”

“…About half of the presented studies make their code publicly available (Cheng et al, 2021;Davari et al, 2021;Heidler et al, 2021;Marochov et al, 2021;Mohajerani et al, 2019;Zhang et al, 2021). Only two-thirds of the described studies make an effort to compare their method with other existing methods by either testing on the same dataset or by retraining and testing the other's model on the own dataset (Cheng et al, 2021;Davari et al, 2022Davari et al, , 2021Hartmann et al, 2021;Heidler et al, 2021;Holzmann et al, 2021;Periyasamy et al, 2022;Zhang et al, 2021). Furthermore, testing on another's dataset alone without retraining one's own model on the other's training dataset beforehand is not a strong comparison, as one's own training dataset might form a better learning basis.…”

“…However, this approach is no longer feasible with the rapidly growing satellite image archives, as manual delineation is a highly time-consuming, tedious, and expensive task. Recent studies (Baumhoer et al, 2019;Cheng et al, 2021;Davari et al, 2022Davari et al, , 2021Hartmann et al, 2021;Heidler et al, 2021;Holzmann et al, 2021;Marochov et al, 2021;Mohajerani et al, 2019;Periyasamy et al, 2022;Zhang et al, 2019Zhang et al, , 2021 have focused on deep learning to extract the calving front and have shown great success. However, the evaluations of these studies have generally been based on different datasets and are, therefore, not comparable.…”

AutoTerm: A "big data" repository of Greenland glacier termini delineated using deep learning

“…This is particularly valuable for the Earth's cryosphere, which exhibits large, non-linear sensitivity to climate change. Recently, several studies have demonstrated that it is possible to use DL methods to delineate glacier termini (Mohajerani et al, 2019;Zhang et al, 2019;Baumhoer et al, 2019;Cheng et al, 2020;Zhang et al, 2021;Davari et al, 2021;Hartmann et al, 2021;Holzmann et al, 2021;Marochov et al, 2021;Davari et al, 2021;Heidler et al, 2021;Periyasamy et al, 2022) with many generating data products that are of interest to the glaciological community.…”