Home

ANR Project n° ANR-12-BS03-003

Unmixing hyperspectral cubes: a big data challenge

The HYPANEMA team addresses the problem of processing large data cubes from hyperspectral imagery. Hyperspectral imaging is undergoing rapid development, but is significantly hampered by the extreme difficulty of extracting relevant information from the great mass of collected data. Within this context, it is therefore necessary to propose a set of specific processing tools that can make the work of scientists easier and more efficient. Analysis and interpretation must generally be preceded by a segmentation step, whose goal is to detect the objects of interest assumed to be spectrally homogeneous. Their spectral and spatial overlap raises the question of how to unmix hyperspectral data – that is, how to estimate the number of pure components or endmembers, their spectral signature and local abundance fractions. Usually, spectral unmixing can be decomposed into two tasks. In a first step of analysis, the spectral signatures of endmembers are estimated. In a second step of inversion, their local abundance fractions are estimated over the entire image.

Machine learning vs. bayesian paradigm

In a supervised mode, which means that the ground truth is available, identification of endmembers can be performed empirically by an operator. In particular, for terrestrial hyperspectral imaging, it is sufficient to define some regions in the image where the presence of these materials is proven. Unmixing is then limited to the inversion step, which leads to the spatial distribution of endmembers with their local abundance. Unfortunately, such a ground truth is not available within the context of astronomical hyperspectral imaging. In this unsupervised case, the analysis of hyperspectral images consists both of identifying endmember spectra and estimating their local abundance fraction from a single image. Like most blind estimation problems, without further assumptions, unsupervised unmixing may not have unique solution. Exploiting constraints inherent in the physical modeling of collected data can then reduce the class of admissible solutions, or even completely eliminate uncertainties. A special case is when the spectrum of an endmember is known, even partially. Unmixing can then consist of extracting this single source.

The literature gives clear evidence that significant progress can be made by using nonlinear models for hyperspectral data unmixing. Despite some methodological difficulties in attempting to analyze and implement them, they would be more likely to reflect the complexity of the underlying physical phenomena. In developing such models for ensuring better fit, taking both spatial and spectral information into account appears very promising. The proposed project aims to develop methods for hyperspectral unmixing based on statistical machine learning framework and Bayesian inference. This class of tools is broad and includes formalisms and algorithms that seem to be particularly appropriate.

Results

The sciences of observation generally contribute to the development of advanced technological equipments. Number of major equipments in the field of the Sciences for the Universe already offer large-field spectrographs, often coupled with an adaptive optics system. This technique should continue to grow with, in the next few years, the deployment of imaging spectrometer MUSE that will equip the VLT. The airborne hyperspectral imagers, or on board of satellites, and dedicated to Earth observation also provide more and more resolved images. Missions are underway or planned by various space agencies: Proba (ESA 2011), EnMAP (Germany, 2014), Prisma (Italy, 2013), HyspIRI (USA, 2013), etc. This technology, with its ability to provide extremely detailed information about the spectral properties of the observed scene, offers numerous opportunities for detection and identification in both civilian and military domains of application.

The development of original algorithms whose functionality and performance are consistent with the high resolution of hyperspectral devices is an issue with high added value. Whatever the application, one of the major problems in analysis and processing of hyperspectral images is that it is extremely difficult to extract relevant information from the mass of generated data. It is therefore necessary to propose to the experimenter, a set of specific tools that can facilitate the work of interpretation regardless of the number of frequency bands of observation.

Perspectives

Processing large hyperspectral cubes is a big data challenge. Thus, skills developed within the context of the HYPANEMA should be easily exploited in other big-data applications, such as radio interferometry with LOFAR and SKA instruments.

News

Kickoff meeting
2013-01-28

The Kickoff meeting for the project was in Nice.

The presentation is available here

Web site creation
2013-01-02

Hypanema website is on!

Recent work

Chen J., Richard C., Honeine P., "Nonlinear Unmixing of Hyperspectral Data Based on a Linear-Mixture/Nonlinear-Fluctuation Model", Signal Processing, IEEE Transactions on, Vol. 61, N. 2, pp 480 -492, 2013.
Abstract: Spectral unmixing is an important issue to analyze remotely sensed hyperspectral data. Although the linear mixture model has obvious practical advantages, there are many situations in which it may not be appropriate and could be advantageously replaced by a nonlinear one. In this paper, we formulate a new kernel-based paradigm that relies on the assumption that the mixing mechanism can be described by a linear mixture of endmember spectra, with additive nonlinear fluctuations defined in a reproducing kernel Hilbert space. This family of models has clear interpretation, and allows to take complex interactions of endmembers into account. Extensive experiment results, with both synthetic and real images, illustrate the generality and effectiveness of this scheme compared with state-of-the-art methods.
BibTeX:
 @article{ chen2013unmixing,
 author = { Chen, J. and Richard, C. and Honeine, P. },
 title = { Nonlinear Unmixing of Hyperspectral Data Based on a Linear-Mixture/Nonlinear-Fluctuation Model }, 
 journal = { Signal Processing, IEEE Transactions on },
 volume = { 61 },
 number = { 2 },
 pages = { 480 -492 },
 year = { 2013 } 
 } 
Altmann Yoann, Dobigeon Nicolas, Tourneret Jean-Yves, "Nonlinearity detection in hyperspectral images using a polynomial post-nonlinear mixing model", IEEE Trans. Image Processing, Vol. 22, N. 4, pp 1267--1276, 2013.
Abstract:
BibTeX:
 @article{ Altmann_IEEE_Trans_IP_2013,
 author = { Altmann, Yoann and Dobigeon, Nicolas and Tourneret, Jean-Yves },
 title = { Nonlinearity detection in hyperspectral images using a polynomial post-nonlinear mixing model }, 
 journal = { IEEE Trans. Image Processing },
 volume = { 22 },
 number = { 4 },
 pages = { 1267--1276 },
 year = { 2013 } 
 } 
Altmann Y., Dobigeon N., McLaughlin S., Tourneret J.-Y., "Nonlinear spectral unmixing of hyperspectral images using Gaussian processes", IEEE Trans. Signal Processing, Vol. 61, N. 10, pp 2442--2453, 2013.
Abstract:
BibTeX:
 @article{ Altmann_IEEE_Trans_SP_2013,
 author = { Altmann, Y. and Dobigeon, N. and McLaughlin, S. and Tourneret, J.-Y. },
 title = { Nonlinear spectral unmixing of hyperspectral images using Gaussian processes }, 
 journal = { IEEE Trans. Signal Processing },
 volume = { 61 },
 number = { 10 },
 pages = { 2442--2453 },
 year = { 2013 } 
 }