Abstract Rock glaciers are an important component of the cryosphere and are one of the most visible manifestations of permafrost. While the significance of rock glacier contribution to streamflow remains uncertain, the contribution is likely to be important for certain parts of the world. High-resolution remote sensing data has permitted the creation of rock glacier inventories for large regions. However, due to the spectral similarity between rock glaciers and the surrounding material, the creation of such inventories is typically conducted based on manual interpretation, which is both time consuming and subjective. Here, we present a novel method that combines deep learning (convolutional neural networks or CNNs) and object-based image analysis (OBIA) into one workflow based on freely available Sentinel-2 optical imagery (10 m spatial resolution), Sentinel-1 interferometric coherence data, and a digital elevation model (DEM). CNNs identify recurring patterns and textures and produce a prediction raster, or heatmap where each pixel indicates the probability that it belongs to a certain class (i.e. rock glacier) or not. By using OBIA we can segment the datasets and classify objects based on their heatmap value as well as morphological and spatial characteristics. We analysed two distinct catchments, the La Laguna catchment in the Chilean semi-arid Andes and the Poiqu catchment in the central Himalaya. In total, our method mapped 108 of the 120 rock glaciers across both catchments with a mean overestimation of 28%. Individual rock glacier polygons howevercontained false positives that are texturally similar, such as debris-flows, avalanche deposits, or fluvial material causing the user's accuracy to be moderate (63.9–68.9%) even if the producer's accuracy was higher (75.0–75.4%). We repeated our method on very-high-resolution Pleiades satellite imagery and a corresponding DEM (at 2 m resolution) for a subset of the Poiqu catchment to ascertain what difference image resolution makes. We found that working at a higher spatial resolution has little influence on the producer's accuracy (an increase of 1.0%), however the rock glaciers delineated were mapped with a greater user's accuracy (increase by 9.1% to 72.0%). By running all the processing within an object-based environment it was possible to both generate the deep learning heatmap and perform post-processing through image segmentation and object reshaping. Given the difficulties in differentiating rock glaciers using image spectra, deep learning combined with OBIA offers a promising method for automating the process of mapping rock glaciers over regional scales and lead to a reduction in the workload required in creating inventories.

中文翻译：

---

使用深度学习和基于对象的图像分析自动检测岩石冰川

摘要 岩石冰川是冰冻圈的重要组成部分，是多年冻土最明显的表现形式之一。虽然岩石冰川对水流贡献的重要性仍不确定，但这种贡献可能对世界某些地区很重要。高分辨率遥感数据允许为大区域创建岩石冰川清单。然而，由于岩石冰川与周围物质之间的光谱相似性，此类清单的创建通常基于人工解释，既费时又主观。这里，我们提出了一种新方法，该方法将深度学习（卷积神经网络或 CNN）和基于对象的图像分析 (OBIA) 结合到一个基于免费提供的 Sentinel-2 光学图像（10 m 空间分辨率）、Sentinel-1 干涉相干数据的工作流程中，以及数字高程模型 (DEM)。CNN 识别重复出现的模式和纹理，并生成预测栅格或热图，其中每个像素表示它属于某个类别（即岩石冰川）的概率。通过使用 OBIA，我们可以根据对象的热图值以及形态和空间特征对数据集进行分割和分类。我们分析了两个不同的流域，智利半干旱安第斯山脉的 La Laguna 流域和喜马拉雅中部的 Poiqu 流域。总共，我们的方法绘制了两个流域 120 个岩石冰川中的 108 个，平均高估了 28%。然而，单个岩石冰川多边形包含结构相似的误报，例如泥石流、雪崩沉积物或河流物质，导致用户的准确度适中 (63.9-68.9%)，即使生产者的准确度更高 (75.0-75.4%) . 我们在非常高分辨率的昴宿星卫星图像和相应的 DEM（分辨率为 2 m）上为 Poiqu 流域的一个子集重复了我们的方法，以确定图像分辨率的差异。我们发现，在更高的空间分辨率下工作对生产者的精度几乎没有影响（增加 1.0%），但是绘制的岩石冰川以更高的用户精度绘制（增加 9.1% 到 72.0%）。通过在基于对象的环境中运行所有处理，可以生成深度学习热图并通过图像分割和对象重塑执行后处理。鉴于使用图像光谱区分岩石冰川的困难，深度学习与 OBIA 相结合，提供了一种很有前景的方法，可以在区域尺度上自动化绘制岩石冰川的过程，并减少创建清单所需的工作量。