 |
| Titel |
Matrix Approximation Techniques for Unsupervised Hyperspectral Data Analysis |
| VerfasserIn |
Albrecht Schmidt, Frédéric Schmidt, Erwan Treguier, Saïd Moussaoui, Nicolas Dobigeon |
| Konferenz |
EGU General Assembly 2010
|
| Medientyp |
Artikel
|
| Sprache |
Englisch
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 12 (2010) |
| Datensatznummer |
250040472
|
|
|
|
|
|
| Zusammenfassung |
| Unsupervised analysis of hyperspectral data delivered by imaging spectrometers
is interesting in many respects in the Planetary Sciences. Since there often
is no ground truth to compare to, unsupervised rather than supervised methods
allow to extract new information from data sets. So far, the practicability
of these methods has suffered from low performance, which makes large-scale
analyses almost prohibitively expensive. New research and implementation
strategies for non-negative matrix factorisation make it possible to extract
sources and relative abundances for typical planetary data sets with
reasonable resources. In this work, we try to give an impression of some of
the trade-offs and opportunities involved.
Non-negative matrix factorisation is a technique which has enjoyed
considerable research and been used in many application areas, from document
clustering to spectral analysis. We compare different approaches in terms of
resource consumption and viability of results in terms of physical
interpretation and meaningfulness in the Planetary Sciences. It turns out
that there are datasets and algorithms for which consistently and efficiently
meaningful results are returned. The results are meaningful in the sense that
abundances as well as extracted source spectra are consistent with community
opinion for well-known examples such as ices on the Martian South Poles. It
is of particular importance to stress that no prerequisite assumptions are
required for the applicability of the algorithms other than the approximate
linearity of combinations of the source spectra. This implies that there are
a great many scenarios in which the algorithms are applicable and deliver
sensible results. To reduce computation times to almost real-time, we
currently sample the input. We also provide estimations of how sampling is
best done for real data sets and try to quantify the trade-offs.
The focus of the presentation is on implementation techniques and comparative
analysis of the performance and limitations of the algorithms. Opportunities
for applications and systems engineering will be discussed. |
|
|
|
|
|
|