National Technical University of Athens (NTUA)
Development of spectral unmixing methods for hypersectral data exploitation
Abstract
dc:description7.4 ΠΕΡΙΛΗΨΗ (ξενόγλωσση)Hyperspectral sensors provide high spectral resolution which contributes to material identification.However, due to low spatial resolution, pixels may contain a mixture of more than one distinctmaterial spectrum. In order to exploit the mixed hyperspectral data, spectral unmixing is applied,which involves the decomposition of the mixed spectra into constituent spectra, also calledendmembers, weighted by their corresponding fractional abundances. This thesis addresses thespectral unmixing problem so that the full potential of hyperspectral data exploitation to be employed.The main contributions include the exploitation of novel concepts and the development of five newmethods dedicated to the steps involved in spectral unmixing; the signal subspace estimation, thedimensionality reduction, the endmember extraction, and the abundance estimation.The first method, called outlier detection method (ODM), is a new automatic non-parametric methodfor estimating the signal subspace dimension. The number of the signal vectors is much lowercompared to the number of the noise vectors. Thus, estimating the population distribution and/or itsstatistical characteristics could comprise errors. The novelty of ODM lies in the fact that it considersonly the existence of noise and treats signals as outliers of noise. It searches for the signals whoseradius is by far larger than the one of the noise introducing for the first time in virtual dimensiontheory, a robust outlier detection method. The ODM achieves the performance of its competitors andoutperforms them in case of small image scenes.The second method is a new band selection (BS) approach for optimizing the performance of theendmember extraction and classification. The few existing BS methods which address spectralunmixing issues set fixed criteria to the spectral information on the whole set of wavelengths. Theyare based on generalised approaches which disregard the spectral characteristics of a particularmaterial of interest, the image diversity and the endmember variability. The proposed methodenables endmember extraction and classification algorithms to act locally in the hyperspectral space.It tries to define subspaces in which spectra of materials vary the most. In case of endmemberextraction, the new method accounts for spectrally closed endmembers (SCEs) – a term introducedin the frame of this study- which are likely overshadowed by the prevalent endmembers, by exploitingthe original bands. In order to extract the SCEs, the concept of multiple convex hulls is used for thefirst time in BS. The proposed approach is effective at detecting low contrast materials, which implydifferent biophysical or chemical properties of a material class. In case of classification, the proposedBS method accounts for subspaces where classes are separable. It achieved the same overallaccuracy as if the whole band set is used, by selecting less than 50% of the total bands.The third method, called simple endmember extraction (SEE) is a new convex-based endmemberextraction method which searches for the most distinct spectra at the vertices of a simplex defined inthe signal subspace. The novelty of the SEE method is that it searches for the extreme values that lieon the end points of the existing transformed axes without further projections that imply iterativeprocedures. Thus, it bears the advantage of simultaneously extracting endmembers with lowcomputational cost. The fourth method is an enhanced version of SEE, called enhanced-SEE (ESEE)and it is an empirical method which compensates the tendency that the majority of the convexbasedendmember extraction methods encounter to select high contrast endmembers over less contrast endmembers. On this account, it changes the distribution of the initial data sample byincreasing the distance between candidate endmembers and the data mean, a technique which hasnever been explored before. Both methods provided the same and even better performancecompared to the state-of-the-arts methods. The computation complexity of the new methods is muchlower than those of vertex component analysis (VCA) and N-FINDR, approximately one and threeorders of magnitude lower, respectively.Last but not least, the fifth method is a new multiple endmember spectral signature analysis(MESMA) based on spectral angle distance, called MESMA-SAD. MESMA accounts for within classspectral variability, however, it needs to calculate all the potential endmember combinations of eachpixel to find the best-fit one, demanding a time-consuming unmixing technology. The new nonparametricmethod significantly minimizes the time-processing compared to the existing MESMAalgorithms by combining the spectral angle distance values and the mean absolute errors.Extensive simulated and real image-based experiments indicate the effectiveness of the proposedmethods to improve the data exploitation, rendering their implementation very promising inhyperspectral image processing.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2014
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Andreou, Charoula
- Ανδρέου, Χαρούλα
Subjects
dc:subject × 13- Υπερφασματική τηλεπισκόπηση
- Φασματικός διαχωρισμός
- Μείωση της διάστασης
- Εξαγωγή καθαρών φασματικών στόχων
- Hyperspectral image
- Remote sensing
- Spectral unmixing
- Dimensionality reduction
- Endmember extraction
- Επιστήμες Μηχανικού και Τεχνολογία
- Άλλες Επιστήμες Μηχανικού και Τεχνολογίες
- Engineering and Technology
- Other Engineering and Technologies
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/37616