Particle filters using sparse observations : applications in palaeoclimatology

Details

Supervisors

Absil, Pierre-Antoine ; Goosse, Hugues

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées

Abstract

The data assimilation problem consists in finding a way to use observations within a model to improve its output. Particle filters are good candidats for data assimilation in palaeoclimatology as model are often non linear. However they often suffer from filter degeneracy as the ensemble of particles may collapse in a small region of the state space. The goal of this work is to present a method based on particle filters that doesn't have this issue. We start from a brief review of existing data assimilation methods and test them on the Lorenz 63 model using twin experiments. Then we apply the sequential importance resampling particle filter and the non linear ensemble transform filter to the LOVECLIM model both in off-line and online twin experiments. Our investigations suggest that the non linear ensemble transform filter is more robust than the sequential importance resampling filter and is able to achieve comparable performances in general.