Abstract Ice cliffs play a key role in the mass balance of debris-covered glaciers, but assessing their importance is limited by a lack of datasets on their distribution and evolution at scales larger than an individual glacier. These datasets are often derived using operator-biased and time-consuming manual delineation approaches, despite the recent emergence of semi-automatic mapping methods. These methods have used elevation or multispectral data, but the varying slope and mixed spectral signal of these dynamic features makes the transferability of these approaches particularly challenging. We develop three semi-automated and objective new approaches, based on the Spectral Curvature and Linear Spectral Unmixing of multispectral images, to map these features at a glacier to regional scale. The transferability of each method is assessed by applying it to three sites in the Himalaya, where debris-covered glaciers are widespread, with varying lithologic, glaciological and climatic settings, and encompassing different periods of the melt season. We develop the new methods keeping in mind the wide range of remote sensing platforms currently in use, and focus in particular on two products: we apply the three approaches at each site to near-contemporaneous atmospherically-corrected Pleiades (2 m resolution) and Sentinel-2 (10 m resolution) images and assess the effects of spatial and spectral resolution on the results. We find that the Spectral Curvature method works best for the high spatial resolution, four band Pleaides images, while a modification of the Linear Spectral Unmixing using the scaling factor of the unmixing is best for the coarser spatial resolution, but additional spectral information of Sentinel-2 products. In both cases ice cliffs are mapped with a Dice coefficient higher than 0.48. Comparison of the Pleiades results with other existing methods shows that the Spectral Curvature approach performs better and is more robust than any other existing automated or semi-automated approaches. Both methods outline a high number of small, sometimes shallow-sloping and thinly debris-covered ice patches that differ from our traditional understanding of cliffs but may have non-negligible impact on the mass balance of debris-covered glaciers. Overall these results pave the way for large scale efforts of ice cliff mapping that can enable inclusion of these features in debris-covered glacier melt models, as well as allow the generation of multiple datasets to study processes of cliff formation, evolution and decline.

中文翻译：

---

使用多光谱卫星图像绘制碎片覆盖的冰川上的冰崖图

摘要 冰崖在被碎片覆盖的冰川的质量平衡中发挥着关键作用，但由于缺乏关于它们在比单个冰川更大的尺度上的分布和演化的数据集，对其重要性的评估受到限制。尽管最近出现了半自动映射方法，但这些数据集通常是使用偏向操作员且耗时的手动描绘方法得出的。这些方法使用了高程或多光谱数据，但这些动态特征的变化斜率和混合光谱信号使得这些方法的可转移性特别具有挑战性。我们开发了三种半自动和客观的新方法，基于多光谱图像的光谱曲率和线性光谱解混，以将冰川上的这些特征映射到区域尺度。通过将其应用于喜马拉雅山的三个地点来评估每种方法的可转移性，在这些地点，碎片覆盖的冰川分布广泛，具有不同的岩性、冰川学和气候环境，并涵盖不同的融化季节时期。我们开发新方法时考虑到当前使用的各种遥感平台，并特别关注两种产品：我们将每个站点的三种方法应用于近同期大气校正的昴宿星团（2 m 分辨率）和哨兵-2（10 m 分辨率）图像并评估空间和光谱分辨率对结果的影响。我们发现光谱曲率方法最适用于高空间分辨率、四波段 Pleaides 图像，而使用解混比例因子对线性光谱解混的修改最适合较粗的空间分辨率，但 Sentinel-2 产品的额外光谱信息。在这两种情况下，冰崖的 Dice 系数都高于 0.48。昴宿星团的结果与其他现有方法的比较表明，光谱曲率方法比任何其他现有的自动或半自动方法性能更好，更稳健。这两种方法都勾勒出大量小的、有时是浅倾斜的、被碎片覆盖的薄冰块，这些冰块与我们对悬崖的传统理解不同，但可能对碎片覆盖的冰川的质量平衡产生不可忽视的影响。