We present a metric learning approach to improve the performance of unsupervised hyperspectral image segmentation. Unsupervised spatial segmentation can assist both user visualization and automatic recognition of surface features. Spatially-continuous segments can improve noise levels and localization of feature boundaries. However, existing segmentation approaches rely on generic measures of spectral similarity. Here we learn domain-specific distance metrics based on training data, improving segment fidelity to semantic categories of interest. Multiclass Linear Discriminant Analysis produces a linear transform that optimally separates a labeled set of training classes. This defines a distance metric that generalizes to new scenes, enabling graph-based segmentations that emphasize key spectral features. We describe tests on data from the Compact Reconnaissance Imaging Spectrometer (CRISM) in which learned metrics improve segment homogeneity with respect to mineralogical categories.Index Terms-Segmentation, Metric Learning, CRISM
HYPERSPECTRAL IMAGE SEGMENTATIONUnsupervised hyperspectral image segmentation can reveal spatial trends that show the physical structure of the scene to an analyst. They highlight borders and reveal areas of homogeneity and change. Segmentations are independently helpful for object recognition, and assist with automated production of symbolic maps. Additionally, a good segmentation can dramatically reduce the number of effective spectra in an image, enabling analyses that would otherwise be computationally prohibitive. Specifically, using an oversegmentation of the image instead of individual pixels can reduce noise and potentially improve the results of statistical post-analysis [1]. Typical segmentation methods for hyperspectral imagery include the watershed transform [2], Markov Random Fields [3], and the Felzenszwalb graph segmentation algorithm [4]. Generally speaking, they cluster pixels based on spatial proximity and a measure of spectral similarity given as a distance metric or other pairwise function. Existing hyperspectral segmentation approaches generally use generic distance measures that treat all channels equally or weight channels based on global statistical properties of the dataset. Such metrics are often confused by noise, instrument artifacts, or spectral variations that are irrelevant to semantic categories of interest.This work aims to improve automatic segmentations by learning a task-relevant measure of spectral distance from expert-labeled training data. We employ a multiclass Linear Discriminant Analysis (LDA) based approach to learn this measure. It produces segmentations that are not only are more visually cohesive, but also quantiatively more accurate in separating known materials into disjoint segments, in comparison to segmentations produced using the Euclidean metric. We evaluate this technique by comparing a set of expert-labeled mineral class maps to the segmentation maps produced by learned metrics, and provide a results on a case study focusing ...