CAlving Fronts and where to Find thEm: a benchmark dataset and methodology for automatic glacier calving front extraction from sar imagery

The temporal variability in calving front positions of marine-terminating glaciers permits inference on the frontal ablation. Frontal ablation, the sum of the calving rate and the melt rate at the terminus, significantly contributes to the mass balance of glaciers. Therefore, the glacier area has been declared as an Essential Climate Variable product by the World Meteorological Organization. The presented dataset provides the necessary information for training deep learning techniques to automate the process of calving front delineation. The dataset includes Synthetic Aperture Radar (SAR) images of seven glaciers distributed around the globe. Five of them are located in Antarctica: Crane, Dinsmoore-Bombardier-Edgeworth, Mapple, Jorum and the Sjörgen-Inlet Glacier. The remaining glaciers are the Jakobshavn Isbrae Glacier in Greenland and the Columbia Glacier in Alaska. Several images were taken for each glacier, forming a time series. The time series lie in the time span between 1995 and 2020. The images have different spatial resolutions, as they were captured by different satellites. The satellites used are Sentinel-1, TerraSAR-X, TanDEM-X, ENVISAT, European Remote Sensing Satellite 1&2, ALOS PALSAR, and RADARSAT-1. Along with the SAR images, two types of labels are provided so that deep learning techniques can be trained in a supervised manner. One label provides the position of the calving front. The other label shows the position of different landscape regions comprising glacier, rock outcrop, ocean including ice-melange, and an area where no information is available consisting of SAR shadows, layover regions, and areas outside the swath. The two labels allow different approaches to calving front delineation, as the calving front can be extracted from landscape region predictions during post-processing. As additional information for post-processing, the dataset includes bounding boxes for the dynamic calving front for each image. This bounding box excludes nearly static calving fronts also visible in the images, which are not of interest but would still be predicted as calving fronts by deep learning techniques. Hence, all front predictions outside this bounding box can be excluded during post-processing. To ensure the generalizability of the trained deep learning techniques to new unseen glaciers, the dataset is split into a training and an out-of-sample test set. The latter shall only be used to test the performance of the trained front delineation algorithm after all hyperparameters are optimized. The test set comprises the time series of Mapple and Columbia. More information on the dataset and how to use it can be found in the related paper.

入海冰川冰崩前缘位置的时间变化特征，可用于反演冰川前缘消融量。冰川前缘消融量即冰崩速率与末端融蚀速率之和，对冰川物质平衡具有显著贡献。因此，世界气象组织（World Meteorological Organization, WMO）已将冰川面积列为关键气候变量（Essential Climate Variable, ECV）产品。本数据集为训练深度学习模型以自动化实现冰崩前缘轮廓提取提供了必要数据支持。本数据集涵盖全球分布的7座冰川的合成孔径雷达（Synthetic Aperture Radar, SAR）影像，其中5座位于南极洲：克兰冰川、丁斯莫尔-邦巴迪尔-埃奇沃思冰川、梅普尔冰川、约鲁姆冰川以及瑟尔根湾冰川；剩余两座分别为格陵兰的雅各布港冰架冰川（Jakobshavn Isbrae Glacier）与阿拉斯加的哥伦比亚冰川。每座冰川均配有多幅时序影像，时间跨度为1995年至2020年。由于影像由不同卫星获取，空间分辨率存在差异，本次使用的卫星包括Sentinel-1、TerraSAR-X、TanDEM-X、ENVISAT、欧洲遥感卫星1&2号、ALOS PALSAR以及RADARSAT-1。除SAR影像外，本数据集还提供两类标签，可用于监督式训练深度学习模型：第一类标签为冰崩前缘的位置标注；第二类标签对不同景观区域进行标注，包括冰川、裸露岩石、包含冰杂冰的海洋区域，以及由SAR阴影、叠掩区域和条带幅宽外区域构成的无有效信息区域。两类标签支持不同的冰崩前缘提取方案：在后期处理阶段，可通过景观区域预测结果间接提取冰崩前缘。为辅助后期处理，本数据集还为每幅影像提供动态冰崩前缘的边界框标注。该边界框会排除影像中几乎静止的冰崩前缘——这类非目标区域虽会被深度学习模型误识别为冰崩前缘，但并非研究关注点。因此在后期处理阶段，可直接排除边界框外的所有冰崩前缘预测结果。为确保训练后的深度学习模型对未见过的冰川具备泛化能力，本数据集被划分为训练集与域外测试集，后者仅用于在所有超参数优化完成后，评估训练好的冰崩前缘提取算法的性能。测试集包含梅普尔冰川与哥伦比亚冰川的时序影像。有关本数据集的更多细节与使用方法，请参阅相关学术论文。