Sie sind hier:



Analysis of hyperspectral image data using machine learning methods

Lehrstuhl: Arbeitsgebiet Bildsignalverarbeitung

Betreuer: Christian Wöhler,

Beginn ab: According to agreement with participants.

Maximale Anzahl der Teilnehmer: 12

Beschreibung: The analysis of hyperspectral image data yields information about the chemical composition of the imaged material. A main application domain is the processing of satellite images. An important method of hyperspectral image analysis is spectral unmixing, which attempts to model an observed spectrum by reconstructing it as a mixture of reference material spectra, the so-called endmembers. These endmember spectra are usually acquired in the laboratory.

Spectral unmixing can be divided into two subproblems: (1) Determination of the endmembers whose spectra are contained in the observed spectrum and (2) computation of the relative fractions of the comprised endmembers. While the second subproblem can commonly be solved based on well-known linear or nonlinear optimisation methods, the first subproblem tends to result in a very high computational complexity, especially when the relevant endmember spectra have to be selected from a large catalogue of possible endmember spectra.

Since modelling of the physical relationships that determine the selection of endmember spectra is complex and often infeasible, it appears to be favourable to employ machine learning methods which are trained based on spectra of materials of known composition. Hence, this project group aims at the development of machine learning methods, especially neural networks and support vector machines, that determine the relevant endmember spectra for an observed spectrum.

The work will be divided into different phases:
- Generation of an appropriate training set (mixture spectra and endmember spectra)
- Definition of an appropriate feature representation of the spectra
- Implementation of neural network classifiers
- Implementation of support vector machine classifiers
- Possibly implementation of further methods, depending on the number of group members
- Implementation of one or more classifiers on FPGA hardware
- Comparison of classifier performances

Figure: Nearly global map of the relative fraction of the mineral pyroxene on the Moon. Background map: LROC WAC mosaic [Speyerer et al., 2011]