Computer-Aided Mapping of Hyper- and Multi-Spectral Data

多光谱和高光谱数据的计算机辅助制图是行星科学中的一个重要应用。从行星表面反射的光在许多波长范围内的许多通道中被获取。这使得能够分析物理和地质表面属性，例如，以便确定某些矿物或岩石的产状，或为未来的空间飞行任务探索行星。为此，多光谱和高光谱图像的大型数据库通常由专家分析，主要以手动方式进行。这个项目的目标是用自动机器学习方法支持这种分析。一大挑战在于缺乏带注释的培训材料，这是从样本数据估计分类器所必需的。为此，该项目提出了遵循主动学习策略的建议。在向专家请求单独的注释之前，以无监督的方式对图像区域进行聚类。这些注释对应于原型区域，并可以传播到类似的和迄今未知的区域。在标注样本数量不断增加的基础上，训练深度神经网络，以进一步自动化映射过程。对自动化决策的不确定性进行建模是项目中最重要的方面之一。机器学习方法的能力变得透明，并提高了专家对结果的可解释性。另一个非常重要的方面是图像区域的表示。区域不是在分类过程中仅分配给单个类别，而是按照属性来表示，所述属性相对于所选属性来表征区域。区域的类别通过它们的属性来识别。这甚至允许识别在训练数据集中看不到的类。这些方法将在多光谱和高光谱图像以及计算机视觉领域考虑的语义分割基准上进行评估。定性分析是由行星地质学家进行的，他评估了这些方法在探索新数据集方面提供的支持。

Computer aided mapping of multi-spectral and hyper-spectral data is an important application in planetary science. The light reflected from a planetary surface is acquired in many channels across a broad range of wavelengths. This allows for analyzing physical and geological surface properties, e.g., in order to determine the occurrence of certain minerals or rocks or to explore planets for future space missions. For this purpose, large databases of multi-spectral and hyper-spectral images are commonly analyzed by experts mostly in a manual fashion. The objective of this project is to support such analyses with automated machine learning methods. A major challenge lies in the lack of annotated training material which is required in order to estimate classifiers from sample data. For this reason, it is proposed in the project to follow an active learning strategy. Image regions are clustered in an unsupervised manner before individual annotations are requested from the expert. The annotations correspond to prototypical regions and can be propagated to similar and so far unknown regions. Based on the increasing amount of annotated samples, deep neural networks are trained in order to automate the mapping process further. Modeling the uncertainty of the automated decisions is one of the most important aspects in the project. The capabilities of the machine learning methods become transparent and improve the interpretability of the results for the expert. Another very important aspect is the representation of image regions. Instead of an assignment to only a single class in the classification process, regions are represented in terms of attributes which characterize a region with respect to selected properties. Classes of regions are recognized by their attributes. This even allows for recognizing classes which are unseen in the training data set. The methods will be evaluated on multi-spectral and hyper-spectral images as well as semantic segmentation benchmarks considered in the computer vision community. A qualitative analysis is performed by a planetary geologist who evaluates the support provided by the methods in the exploration of a new data set.