TY  - GEN
N2  - Environmental systems are
nonlinear, multiscale and non-separable. Mathematical models describing these systems
are typically high-dimensional and always have missing physics. Therefore, determining
the system?s state and its future development relies on in situ observations. Information
from models and observations are combined using data assimilation methods, which are
mainly developed for divergent systems as they arise from weather prediction. Applying
them also to convergent systems requires modifications of these methods. I investigated
the differences of data assimilation in convergent and divergent systems and found that
parameter estimation is essential in convergent systems. In this work, I enhanced the
particle filter, an ensemble-based data assimilation method. In contrast to other methods,
the particle filter is able to handle nonlinear systems and to describe the resulting
non-Gaussian probability density functions. However, for parameter estimation modifications
of the resampling, i.e. the renewal of the ensemble, are necessary. I developed a
resampling method that uses the weighted covariance information calculated from the
ensemble to generate new particles. This method correlates observed with unobserved
dimensions and can effectively estimate state and parameters in a convergent system.
To be applicable in high-dimensional systems, particle filters need localisation. The
introduced resampling allows localisation, which further increases the efficiency of the
filter.
A1  - Berg, Daniel
AV  - public
Y1  - 2018///
UR  - https://archiv.ub.uni-heidelberg.de/volltextserver/25182/
ID  - heidok25182
TI  - Particle Filters for
Nonlinear Data Assimilation
ER  -