A Labeled Image Dataset for Deep Learning-Driven Rockfall Detection on the Moon and Mars

Background: The term rockfall describes the rapid displacement of a large, usually meter-sized block of rock down-slope, triggered by, for example, endogenic or exogenic events like impacts, quakes or rainfall. In a remote sensing context, the term rockfall is also being used to describe the characteristic geomorphic deposit of a rockfall event that can be identified from an air- or space-borne perspective, i.e., the combination of a displaced boulder and the track it carved into the slope substrate while bouncing, rolling, and sliding over the surface (also called `boulder with track' or `rolling boulder'). In planetary science, the spatial distribution and frequency of rockfalls provide insights into the global erosional state and activity of a planetary body while their tracks act as tools that allow for the remote estimation of the surface strength properties of yet unexplored regions in preparation of future ground exploration missions, such as the lunar pyroclastic, polar sunlit and permanently shadowed regions of the Moon. Due to their small physical size (meters), the identification and mapping of rockfalls in planetary satellite imagery is challenging and very time-consuming, however. For this reason, Bickel et al. (2018) and Bickel et al. (2020) trained convolutional neural networks to automate rockfall mapping in lunar and martian satellite imagery. Parts of the unpublished datasets used for earlier work have now been complemented with newly labeled data to create a well-balanced dataset of 2,822 lunar and martian rockfall labels (which we call `RMaM-2020' --- [R]ockfall [Ma]rs [M]oon [2020], 416 MB in total, available here) that can be used for deep learning and other data science applications. Here, balanced means that the labels have been derived from imagery with a wide and continuous range of properties like spatial resolution, solar illumination, and others. So far, this dataset has been used to analyze the benefits of multi-domain learning on rockfall detector performance (Mars & Moon vs. Moon-only or Mars-only), but there are numerous other (non-planetary science) applications such as for featurization, feature or target recognition (aircraft/spacecraft autonomy), and data augmentation experiments.

背景：岩崩（rockfall）指大型、通常为米级的岩块沿坡向下快速位移，其触发因素包括内动力或外动力事件，例如撞击、地震或降雨。在遥感场景中，岩崩一词也可指代从航空或航天视角可识别的岩崩事件特有的地貌堆积体，即位移岩块与其在坡面基质上弹跳、滚动、滑动时刻划出的运动轨迹的组合（也称为“带轨迹岩块”或“滚动岩块”）。在行星科学领域中，岩崩的空间分布与频率可为行星天体的全球侵蚀状态与活动程度提供研究线索，而其运动轨迹则可作为工具，用于遥感估算尚未探索区域的表面强度特性，为未来地面探测任务（例如月球火山碎屑沉积区、极地日照区及永久阴影区）提供前期参考。不过由于岩崩的物理尺寸较小（仅米级），在行星卫星影像中识别并绘制岩崩分布极具挑战性，且耗时极长。为此，Bickel等人（2018）与Bickel等人（2020）训练了卷积神经网络（convolutional neural networks），以实现月球与火星卫星影像中岩崩分布的自动化制图。此前研究中使用的部分未公开数据集，现已补充新标注数据，最终构建出包含2822条月球与火星岩崩标注的平衡数据集（我们将其命名为“RMaM-2020”——即[R]ockfall [Ma]rs [M]oon [2020]，总大小416MB，可于此处获取），可用于深度学习及其他数据科学相关研究。此处的“平衡”指该数据集的标注样本源自空间分辨率、太阳光照条件等属性覆盖范围广泛且连续的影像数据。截至目前，该数据集已被用于分析多域学习对岩崩检测模型性能的增益（例如火星与月球联合训练 vs. 仅月球或仅火星训练），同时还可应用于诸多其他（非行星科学领域）场景，例如特征工程、特征或目标识别（航空器/航天器自主系统）以及数据增强实验。