Hii - HAT : an IDL / ENVI Toolkit for Rapid Hyperspectral Inquiry

Introduction: Significant challenges face any user of planetary hyperspectral imagery. For the scientist, the sheer volume of data precludes exhaustive manual analysis. For the mission planner, the complex subtleties of hyperspectral images are difficult to summarize in a readily interpretable product. Both experts would benefit from a rapid, robust means to create draft mineralogical maps, summarize novel detections, and generally draw attention to areas of interest for further investigation. In this work we discuss the Hii-HAT (Hyperspectral Image Interactive Helper and Analysis Tools) toolset developed for the IDL/ENVI environment. It attends to the specific requirements of planetary geologists, such as low signal-to-noise ratios and a lack of reference spectra from the surface. Hii-HAT incorporates several novel algorithms, including the concept of superpixel decomposition for noise removal and image feature enhancement. It assists in the discovery of endmembers, forms mineral maps through interactive unmixing, and assists in the detection of appropriate neutral regions for a given region of interest (ROI). Here we explore its capabilities with imagery from the Compact Reconnaissance Imaging Spectrometer (CRISM) instrument orbiting Mars [1], specifically the 1000-2500nm wavelengths of images (frt0000)3e12, 8158, 863e, and 3fb9. Superpixels: Manual analysis often focuses on either an individual spectrum at a single pixel or on the mean spectrum of a large ROI. The former approach preserves spatial resolution but is sensitive to measurement noise. The later reduces noise but requires laborious manual segmentation. Automating segmentation, therefore, is of great interest with the caveat that any errors can easily average-out interesting signals. Many of Hii-HAT’s functions exploit a superpixel representation that combines benefits from both approaches. Superpixels represent the image as contiguous regions a few tens or hundreds of pixels in area. By erring on the side of oversegmentation, superpixels reduce noise while preserving small signals evident in only a few contiguous image pixels. This preprocessing step can improve further spectral analysis by simply replacing individual pixel spectra by the mean superpixel spectra, yielding three main advantages. First, measurement noise is reduced proportionally to the square root of the superpixel area. Second, the superpixel’s boundary can help discern subtle hyperspectral features in otherwise bland areas. Third, reducing the number of spectra required for processing (in our case by 100 times) speeds successive algorithms and enables new classes of automated analysis. Superpixels exploit the fact that physical surface features are spatially contiguous, and spatial constraints identify populations of pixels drawn from the same feature. Many segmentation strategies might produce reasonable superpixel segmentations. Hii-HAT utilizes the Felzenszwalb graph segmentation method for its computational efficiency and the ability to accommodate any spectral distance metric [2]. Endmember Extraction: Hii-HAT generates several automatic summary products including superpixelaugmented endmember extraction. By the geographic mixing assumption, observed reflectances are linear combinations of several pure endmember materials. These are of great interest as they represent the physically purest minerals in a scene – the archetypes and novelties that drive exploration. A noise-reduced superpixel representation can improve the performance of classical endmember detection algorithms. Contiguous spatial regions in an image are likely to contain similar mineralogy lending a physical interpretation to the shape of the endmember superpixel.