NOAH-H, a deep-learning, terrain classification system for Mars: Results for the ExoMars Rover candidate landing sites
NOAH-H, a deep-learning, terrain classification system for Mars: Results for the ExoMars Rover candidate landing sites
复制标题
DOI：
10.1016/j.icarus.2021.114701
复制
发表时间：
2022-01
期刊：
影响因子：
3.2
通讯作者：
A. Barrett;M. Balme;M. Woods;S. Karachalios;Danilo Petrocelli;L. Joudrier;E. Sefton-Nash
中科院分区：
文献类型：
--
作者：
A. Barrett;M. Balme;M. Woods;S. Karachalios;Danilo Petrocelli;L. Joudrier;E. Sefton-Nash
在这项调查中，深度学习地形分类系统，“新奇或异常猎人- HiRISE”（NOAH-H），用于分类高分辨率成像科学实验（HiRISE）图像的Oxia Planum和Mawrth Vallis。一套本体类的开发，涵盖了各种表面纹理和风成底形目前在这两个网站。这些类的标记类型的例子被用来训练一个深度神经网络（DNN）进行语义分割，以确定这些类在进一步的HiRISE images.This贡献讨论的方法和结果的研究从地貌学家的角度来看，提供了一个案例研究应用机器学习的景观分类任务。我们的目标是强调如何编译训练数据集，选择本体类，并了解这些系统可以做什么和不能做什么。我们强调的问题时出现的适应传统的行星映射工作流程的生产训练数据。我们讨论了模型的像素尺度精度，以及定性因素如何影响输出的可靠性和可用性。我们得出结论，“景观级”的可靠性是人类使用输出栅格的关键。输出结果通常比像素级精度统计数据更有用，但必须谨慎对待产品，而不是将其视为地质起源的最终仲裁者。很好地理解模型如何以及为什么对不同的景观特征进行分类，对于可靠地解释它至关重要。当使用得当的分类栅格提供了一个很好的指示不同的地形类型的流行和分布，并告知我们的研究领域的理解。因此，我们的结论是，它是适合的目的，并适合用于进一步的工作。
In this investigation a deep learning terrain classification system, the “Novelty or Anomaly Hunter – HiRISE” (NOAH-H), was used to classify High Resolution Imaging Science Experiment (HiRISE) images of Oxia Planum and Mawrth Vallis. A set of ontological classes was developed that covered the variety of surface textures and aeolian bedforms present at both sites. Labelled type-examples of these classes were used to train a Deep Neural Network (DNN) to perform semantic segmentation in order to identify these classes in further HiRISE images.This contribution discusses the methods and results of the study from a geomorphologists perspective, providing a case study applying machine learning to a landscape classification task. Our aim is to highlight considerations about how to compile training datasets, select ontological classes, and understand what such systems can and cannot do. We highlight issues that arise when adapting a traditional planetary mapping workflow to the production of training data. We discuss both the pixel scale accuracy of the model, and how qualitative factors can influence the reliability and usability of the output.We conclude that “landscape level” reliability is critical for the use of the output raster by humans. The output can often be more useful than pixel scale accuracy statistics would suggest, however the product must be treated with caution, and not considered a final arbiter of geological origin. A good understanding of how and why the model classifies different landscape features is vital to interpreting it reliably. When used appropriately the classified raster provides a good indication of the prevalence and distribution of different terrain types, and informs our understanding of the study areas. We thus conclude that it is fit for purpose, and suitable for use in further work.