Superpixel segmentation for analysis of hyperspectral data sets, with application to Compact Reconnaissance Imaging Spectrometer for Mars data, Moon Mineralogy Mapper data, and Ariadnes Chaos, Mars

Abstract

[1] We present a semiautomated method to extract spectral end-members from hyperspectral images. This method employs superpixels, which are spectrally homogeneous regions of spatially contiguous pixels. The superpixel segmentation is combined with an unsupervised end-member extraction algorithm. Superpixel segmentation can complement per pixel classification techniques by reducing both scene-specific noise and computational complexity. The end-member extraction step explores the entire spectrum, recognizes target mineralogies within spectral mixtures, and enhances the discovery of unanticipated spectral classes. The method is applied to Compact Reconnaissance Imaging Spectrometer for Mars (CRISM) images and compared to a manual expert classification and to state-of the-art image analysis techniques. The technique successfully recognizes all classes identified by the expert, producing spectral end-members that match well to target classes. Application of the technique to CRISM multispectral data and Moon Mineralogy Mapper (M3) hyperspectral data demonstrates the flexibility of the method in the analysis of a range of data sets. The technique is then used to analyze CRISM data in Ariadnes Chaos, Mars, and recognizes both phyllosilicates and sulfates in the chaos mounds. These aqueous deposits likely reflect changing environmental conditions during the Late Noachian/Early Hesperian. This semiautomated focus-of-attention tool will facilitate the identification of materials of interest on planetary surfaces whose constituents are unknown.