Interactive visual analysis and classification of hyperspectral imaging data

Abstract

Hyperspectral imaging is a widely used remote sensing technique in planetary sciences. Captured data consist of arrays of images of the same scene taken at a high number of sensor wavelengths. Studying these data, scientists search to understand, for example, the mineral composition of the surface, or types and kinds of vegetation present in the region. Different mineral or vegetation classes induce spectral features that allow for their identification and mapping. The two main tasks are selecting a subset of bands that best capture the relevant spectral features and separating different mixtures from summary products in these bands. It is common practice that the subsets of bands used in the analysis are predefined based on experience, i.e., without taking into account actual data. Then, classification of the regions is performed by analyzing the summary products. We, instead, propose an approach that allows for data-driven selection of subsets of bands and for a more accurate separation of different mixtures. Our approach relies on an interactive visual analysis using suitable visual encodings and interaction mechanisms. In a first step, we produce a similarity plot of the bands of the hyperspectral imaging data by employing a projection-based dimensionality reduction technique. The similarity plot allows for the selection of most informative bands. In a second step, we apply an automatic hierarchical density-based clustering approach to the pixels of the selected bands. The resulting cluster hierarchy is interactively explored and adjusted using coordinated views of a cluster tree visualization, a parallel coordinate plot of the bands, and a spatial data visualization. Brushing and linking in the coordinated views allows for an intuitive interactive analysis of the bands leading to the desired mineral mapping result. The effectiveness of our approach is demonstrated by applying it to mineral mapping of Compact Reconnaissance Imaging Spectrometer for Mars data and vegetation classification of Airborne Visible/Infrared Imaging Spectrometer data.