|
Title
|
|
|
|
Development of advanced hyperspectral unmixing methods
| |
|
Author
|
|
|
|
| |
|
Abstract
|
|
|
|
| Hyperspectral cameras collect the reflected light of materials in hundreds of narrow, contiguous spectral bands in the visible, near and shortwave infrared wavelengths to provide a continuous reflectance spectrum for each pixel. Due to the complex interaction of light with materials, these spectra are highly nonlinear mixtures of the reflectances of the material constituents. The general goal of this thesis is to estimate the composition of materials from reflectance spectra. Mixing models describe the reflectance spectrum of a material as a (nonlinear) mixture of the constituent materials. The main disadvantage of these models is that the model parameters are not properly interpretable in terms of the fractions. Moreover, not all spectra necessarily follow the same particular mixing model. Alternatively, the complex mixing effects can be learned using supervised machine learning methods. This requires ground truth training data, in the form of the actual compositions (i.e., the spectra and fractions of the constituents). One major drawback of these strategies is that the estimated fractions do not comply with their physical constraints, leading to a loss of the physical meaning of the estimated parameters. Another disadvantage of the learned models is that they cannot perform well in case training and test spectra are obtained under different environmental conditions or by different sensors, causing spectral variability of the acquired spectra. In this thesis, a hybrid framework was developed that combines the physical interpretability of a model and the flexibility of data-driven approaches. The general idea is to learn the complex relation between the nonlinear spectra and spectra that follow a particular mixing model by utilizing advanced machine learning regression algorithms. Based on this strategy, a number of different nonlinear unmixing methods were developed: 1) A supervised method that learns a mapping between the nonlinear spectra and the linear mixing model. 2) A strategy for the estimation of leaf biochemical parameters from leaf reflectance and transmittance spectra, by learning a mapping to a leaf biochemical model (PROSPECT). 3) A semi-supervised method to reduce the number of training samples required to learn the nonlinearities, and additionally does not require the availability of pure pixels. 4) A robust supervised method for nonlinear spectral unmixing that is invariant to endmember variability. |
| |
|
Language
|
|
|
|
English
| |
|
Publication
|
|
|
|
Antwerpen
:
Universiteit Antwerpen, Faculteit Wetenschappen, Departement Fysica
,
2020
| |
|
Volume/pages
|
|
|
|
139 p.
| |
|
Note
|
|
|
|
:
Scheunders, Paul [Supervisor]
| |
|
Full text (open access)
|
|
|
|
| |
|