Sparse Superpixel Unmixing of CRISM Hyperspectral Images 1 NASA / Caltech / JPL / Instrument Software and Science Data Systems Images courtesy NASA / Caltech JPL / Brown University. This presentation Copyright 2009 California Institute of Technology. US Government Support Acknowledged. David R. Thompson, JPL Martha S. Gilmore, Wesleyan University Becky Castaño, JPL
Sparse Superpixel Unmixing Problem Background Sparse Unmixing Superpixel Segmentation Preliminary Results 2 NASA / Calech / JPL / Instrument Software and Science Data Systems Agenda MRO (Courtesy NASA/JPL/Caltech)
Motivation 3 NASA / Caltech / JPL / Instrument Software and Science Data Systems
Motivation “Intelligent Assistant” for data mining, fast image analysis Tactical observation selection Detection of anomalous or important mineralogy Challenges: Source constituents unknown High signal to noise Sparse unmixing Recovers constituents from an overcomplete source library Superpixel segmentation speeds results for whole images NASA / Caltech / JPL / Instrument Software and Science Data Systems 4 multispectral (survey) hyperspectral (targeted)
Sparse unmixing Unmixing with an overcomplete source library Linear mixing model NASA / Calech / JPL / Instrument Software and Science Data Systems 5 Mixing coefficients Overcomplete library of source signals Gaussian noise Reconstruction Constituents Phyllosilicate Mafics
Bayesian Unmixing Sparsity-inducing exponential prior on mixing coefficients Objective function: maximize p(coefficients|data) Gradient ascent [similar to Moussaui et al. 2008] NASA / Calech / JPL / Instrument Software and Science Data Systems 6 Controls sparsity
Datasets and Preprocessing Compact Reconnaissance Imaging Spectrometer (CRISM) images of Nili Fossae region “Full-resolution targeted” images frt00003e12, frt00003fb9 (233 bands in 1.0 to 2.5 micrometer range) Atmospheric correction with Volcano division NASA / Calech / JPL / Instrument Software and Science Data Systems 7 frt00003e12 frt00003fb9
Bayesian Unmixing NASA / Calech / JPL / Instrument Software and Science Data Systems 8 Constituents Site B reconstruction Constituents Mafics Site A reconstruction Phyllosilicate Mafics
MCMC Probabilistic Unmixing Gibbs sampler for mixing coefficients, proposal distributions based on multivariate Gaussian NASA / Calech / JPL / Instrument Software and Science Data Systems 9
Sparse Superpixel Unmixing Problem Background Datasets & Preprocessing Sparse Unmixing Superpixel Segmentation Preliminary Results 10 NASA / Calech / JPL / Instrument Software and Science Data Systems Agenda MRO (Courtesy NASA/JPL/Caltech)
Superpixel Segmentation “Superpixels” are image segments corresponding to homogeneous sub-regions [Ren et al. 2003, Mori et al 2005] Potential advantages: Noise reduction Faster processing NASA / Calech / JPL / Instrument Software and Science Data Systems 11 Image created with code by Mori et al., Courtesy CMU
Superpixel Segmentation Felzenszwalb graph partitioning algorithm [Felzenszwalb et al. 2004] Compute edge weights using Euclidean distance between spectra NASA / Calech / JPL / Instrument Software and Science Data Systems 12
Superpixel Segmentation Felzenszwalb graph partitioning algorithm [Felzenszwalb et al. 2004] NASA / Calech / JPL / Instrument Software and Science Data Systems 13 Iteratively join segments when there is no evidence of a boundary between them
Felzenszwalb graph partitioning algorithm [Felzenszwalb et al. 2004] Superpixel Segmentation Compare strongest joining edge to weakest edge of spanning trees Weighted with an additive bias prevents small regions NASA / Calech / JPL / Instrument Software and Science Data Systems 14
15 NASA / Calech / JPL / Instrument Software and Science Data Systems Superpixel Segmentation originalcoarsefine
Mapping Results Abundance measure produced by combining mixing coefficients from Olivine, Phyllosilicate library samples Evaluated correlation with hand-crafted summary products NASA / Calech / JPL / Instrument Software and Science Data Systems 16 Olivine detections OLINDEX standard Phyllosilicate detections D2300 standard
Mapping Results High correlation scores for both minerals, images NASA / Calech / JPL / Instrument Software and Science Data Systems 17 ImageIndexSegment- ation Corr.Precis.Recall 3e12OLINDCoarse Fine D2300Coarse Fine fb8OLINDCoarse Fine
Conclusions Superpixel segmentation has utility for fast summary data products Demonstration of gradient ascent unmixing with sparsity-inducing priors NASA / Calech / JPL / Instrument Software and Science Data Systems 18 MRO (Courtesy NASA/JPL/Caltech)
Future Work Superpixel-enhanced endmember extraction NASA / Calech / JPL / Instrument Software and Science Data Systems 19 Traditional endmember extraction, SMACC algorithm (noise artifacts, 3/5 actual classes detected) New automatic method based on superpixels (5/5 actual classes detected) “Ground truth” classes from geologist classification
Future Work Superpixel-enhanced endmember extraction Endmember superpixels serve as regions of interest for automated feature detection NASA / Calech / JPL / Instrument Software and Science Data Systems 20 Mean spectrum of target region
MCMC Probabilistic Unmixing 21 NASA / Calech / JPL / Instrument Software and Science Data Systems
Acknowledgements Thanks to Brown University for the CAT/ENVI tools used in atmospheric correction and reprojection Sponsorship by NASA AMMOS / MGSS Multimission Ground Support hyperspectral.jpl.nasa.gov NASA / Calech / JPL / Instrument Software and Science Data Systems 22
Backup Slides 23 NASA / Calech / JPL / Instrument Software and Science Data Systems
Superpixel Segmentation Felzenszwalb graph partitioning algorithm [Felzenszwalb et al. 2004] Merge contiguous subregions using Euclidean distance between spectra NASA / Calech / JPL / Instrument Software and Science Data Systems 24 ?
Superpixel Segmentation Felzenszwalb graph partitioning algorithm [Felzenszwalb et al. 2004] Merge contiguous subregions using Euclidean distance between spectra NASA / Calech / JPL / Instrument Software and Science Data Systems 25 ?
Superpixel Segmentation Felzenszwalb graph partitioning algorithm [Felzenszwalb et al. 2004] Merge contiguous subregions using Euclidean distance between spectra NASA / Calech / JPL / Instrument Software and Science Data Systems 26
1. Sparse unmixing discovers constituents from an overcomplete source library 1. Draft mineralogical maps Motivation NASA / Caltech / JPL / Instrument Software and Science Data Systems 27 Reconstruction Constituents Phyllosilicate Mafics Phyllosilicate detections